publications from the thesis - shodhgangashodhganga.inflibnet.ac.in/bitstream/10603/13127/17/17...2....
TRANSCRIPT
Publications from the Thesis
Papers in International Journals
1. Syed Atiqur Rahman and Mohd. Samar Ansari. A Neural
Circuit with Transcendental Energy Function for Solving
System of Linear Equations. Analog Integrated Circuits and
Signal Processing, 66(3):433–440, 2011.
2. Mohd. Samar Ansari and Syed Atiqur Rahman. DVCC-
Based Non-linear Feedback Neural Circuit for Solving Sys-
tem of Linear Equations. Circuits Systems and Signal Process-
ing, 30(5):1029–1045, 2011.
3. Mohd. Samar Ansari and Syed Atiqur Rahman. Non-
linear Feedback Neural Network for Solution of Quadratic
Programming Problems. International Journal of Computer
Applications, 39(2):44–48, 2012.
4. Mohd. Samar Ansari and Syed Atiqur Rahman. A Non-
Linear Feedback Neural Network for Graph Coloring. In-
ternational Journal of Computer Applications, 39(16):31–33, 2012.
5. Mohd. Samar Ansari and Syed Atiqur Rahman. A Non-
Linear Feedback Neural Network for Solving Quadratic
Programming Problems. Journal of Active and Passive Elec-
tronic Devices, Accepted for Publication. Reference ID: RC09-53.
Book Chapter
1. S. Maheshwari, Mohd. Samar Ansari and S. A. Rahman.
Linear and Non-Linear Applications of CMOS DVCC. In
CMOS Technology, Nova Science Publishers, U.S.A., 2010.
Papers in International Conferences
1. S.A. Rahman and M.S. Ansari. Solution of Simultaneous
Linear Equations using Feedback Neural Network. In Pro-
ceedings of International Conference on Recent Applications of Soft
Computing in Engineering and Technology (RASIET-07), IET, Al-
war, India, 2007.
2. S.J. Arif, S.A. Rahman and M.S. Ansari. A NonLinear
Feedback Neural Network for Graph Coloring. In Pro-
ceedings of International Conference on Recent Applications of Soft
Computing in Engineering and Technology (RASIET-07), IET, Al-
war, India, 2007.
3. S.A. Rahman and M.S. Ansari. Solution of Quadratic Pro-
gramming Problem using Feedback Neural Network. In
Proceedings of International Conference on Intelligent Systems and
Networks (ISN-2008), Computational Intelligence Laboratory, In-
stitute of Science and Technology, Klawad, India, 2008.
4. Mohd. Samar Ansari and Syed Atiqur Rahman. A non-
linear neural circuit for solving system of simultaneous
linear equations. In Proceedings of International Conference
on Multimedia, Signal Processing and Communication Technolo-
gies (IMPACT-2009), Aligarh, 2009.
5. Mohd. Samar Ansari and Syed Atiqur Rahman. A novel
current-mode non-linear feedback neural circuit for solv-
ing linear equations. In Proceedings of International Conference
on Multimedia, Signal Processing and Communication Technologies
(IMPACT-2009), Aligarh, 2009.
6. Mohd. Samar Ansari. QPP: Solution using feedback neu-
ral network employing transconducting synaptic intercon-
nections. In Proceedings of International Conference on Intelli-
gent Systems and Nanotechnology (IISN-2010), Computational In-
telligence Laboratory, Institute of Science and Technology, Klawad,
India, 2010.
7. Mohd. Samar Ansari and Syed Atiqur Rahman. A Voltage-
Mode Analog Circuit for Solving Linear Programming Prob-
lems. In th International Conference on Computer Applications
in Electrical Engineering Recent Advances (CERA-09), IIT Roor-
kee, India, 2010.
8. Mohd. Samar Ansari and Syed Atiqur Rahman. DVCC-
based Non-Linear Feedback Neural Circuit for Solving
Quadratic Programming Problems. In th International Con-
ference on Computer Applications in Electrical Engineering Recent
Advances (CERA-09), IIT Roorkee, India, 2010.
9. Mohd. Samar Ansari and Syed Atiqur Rahman. A DVCC-
based Non-Linear Analog Circuit for solving Linear Pro-
gramming Problems. In International Conference on on Power,
Control and Embedded Systems (ICPCES-2010), MNNIT, Allahabad,
2010.
10. Mohd. Samar Ansari. Employing Differential Voltage
Current Conveyor in Graph Coloring Applications. In
International Conference on on Power, Control and Embedded Sys-
tems (ICPCES-2010), MNNIT, Allahabad, 2010.
Papers Under Review
1. Mohd. Samar Ansari and Syed Atiqur Rahman. Current-
mode Integrable Non-Linear Feedback Neural Circuit for
Solving Linear Equations. Circuits Systems and Signal Pro-
cessing
.
2. Mohd. Samar Ansari and Syed Atiqur Rahman. Current-
mode Integrable Non-Linear Feedback Neural Circuit for
Solving Linear Equations. Radioengineering
.
A neural circuit with transcendental energy function for solvingsystem of linear equations
Syed Atiqur Rahman • Mohd. Samar Ansari
Received: 16 November 2008 / Revised: 16 March 2010 / Accepted: 6 August 2010 / Published online: 21 August 2010
Springer Science+Business Media, LLC 2010
Abstract A feedback neural network based circuit to
solve a system of simultaneous linear equations is pre-
sented. The circuit has an associated transcendental energy
function that ensures fast convergence to the exact solution
while enjoying reduction in hardware complexity over
existing schemes. The proof of the energy function has
been given and it is shown that the gradient network con-
verges exactly to the solution of the system of equations.
PSPICE simulation results are presented for linear systems
of equations of various sizes and are found to agree exactly
with the algebraic solution. Hardware implementation for
small sized problems further confirm the circuit operation.
Keywords Neural network applications Neural network
hardware Nonlinear circuits Linear algebra Linear equations
1 Introduction
The solution of linear system of equations is key to many
real-time engineering applications viz. real-time speech
coding, image processing, stochastic modelling, and com-
puter-aided realistic three-dimensional image synthesis
[1–4]. The problems to be solved are often of very large
size, so that large computer resources or massive parallel
computer systems with distributed memory are needed to
solve them [1]. Massively parallel processing and fast
convergence, which are inherent to neural networks, are
utilized to reduce the time taken to arrive at the solution
[1–3]. However, existing neural networks have variable
penalty parameters which decrease to zero as time
increases to infinity in order to get better accuracy of
solution. Therefore, the time taken to arrive at the solution
in such cases is not always short [3, 4].
Hardware solutions based on neural networks have been
put forward by Xia et al. [2], Cichocki and Ubehauen [3]
and Wang [4]. To solve an n-variable system of equations,
the network of [4] uses three operational amplifiers, one
capacitor and (n ? 5) resistances to emulate a single neu-
ron and the time to arrive at the solution is of the order of
hundreds of milliseconds. The network of [3] provides a
significantly improved solution time of around a micro-
second but at the cost of increased hardware complexity.
Each neuron in [3] comprises of three weighted summers
and an inverting integrator. The architecture of [2], which
is a generalized neural network based implementation of
Censor and Elfving’s method for linear inequalities, also
uses an approach similar to [4] utilizing weighted adders
and integrators to realize the neurons.
A new architecture with fast convergence and reduced
circuit complexity is presented. The proposed architecture
uses non-linear feedback which leads to a new energy
function that involves transcendental terms. This tran-
scendental energy function is fundamentally different from
the standard quadratic form associated with Hopfield net-
work and its variations. Along with presenting the analysis
of the proposed circuit and the proof of the energy func-
tion, it is also shown that the stable state of the network
corresponds exactly with the solution of the given system
of linear equations.
S. A. Rahman M. S. Ansari (&)
Department of Electronics Engineering, Aligarh Muslim
University, Aligarh, India
e-mail: [email protected]
S. A. Rahman
e-mail: [email protected]
123
Analog Integr Circ Sig Process (2011) 66:433–440
DOI 10.1007/s10470-010-9524-2
2 Proposed neural network
Let the simultaneous linear equations to be solved are
AV ¼ B ð1Þ
where
A ¼
a11 a12 . . . a1n
a21 a22 . . . a2n
..
. ...
. . . ...
an1 an2 . . . ann
26664
37775 ð2Þ
B ¼
b1
b2
..
.
bn
26664
37775 ð3Þ
V ¼
V1
V2
..
.
Vn
26664
37775 ð4Þ
where V1, V2,…, Vn are the variables and aij and bi are
constants. We will assume that the coefficient matrix A is
invertible, and hence, the system of linear equations (1) is
consistent and not under-determined. In other words, the
linear system (1) has a uniquely determined solution.
The proposed neural-network based circuit to solve the
system of equations of (1) is presented in Fig. 1. As can be
seen from Fig. 1, individual equations from the set of
equations to be solved are passed through non-linear syn-
apses which are realized using comparators.
The output of the i-th comparator can be modelled by
xi ¼ Vm tanh b ai1V1 þ ai2V2 þ þ ainVn bið Þ ð5Þ
where b is the open-loop gain of the comparator (practi-
cally very high), ± Vm are the output voltage levels of the
comparator and V1, V2,…, Vn are the neuron outputs.
The outputs of the comparators are fed to neurons
having weighted inputs. These weighted neurons are real-
ized by using opamps where the resistors Rij act as weights.
Rpi and Cpi are the input resistance and capacitance of the
opamp corresponding to the i-th neuron. These parasitic
components are included to model the dynamic nature of
the opamp.
Node equation for node ‘Ni’ gives the equation of
motion of the i-th neuron as
Cpidui
dt¼ x1
R1iþ x2
R2iþ þ xn
Rni ui
1
Ri
ð6Þ
where ui is the internal state of the i-th neuron, and
Ri ¼ R1i k R2i k Rni k Rpi ð7Þ
Using (5) in (6) results in
Cpidui
dt¼ Vm tanh b a11V1 þ a12V2 þ þ a1nVn b1ð Þ
R1i
þ Vm tanh b a21V1 þ a22V2 þ þ a2nVn b2ð ÞR2i
þ þ Vm tanh b an1V1 þ an2V2 þ þ annVn bnð ÞRni
ui
Rið8Þ
Moreover, as has been shown in section-III, the network
in Fig. 1 can be associated with an Energy Function ‘E’
given by
E ¼ Vm
Xn
i¼1
ln cosh bXn
j¼1
aijVj bi
!Xn
i¼1
1
Ri
ZVi
0
uidV
ð9Þ
From (9), it follows that
oE
oV1
¼Vma11 tanh b a11V1 þ a12V2 þ þ a1nVn b1ð Þ
þ Vma21 tanh b a21V1 þ a22V2 þ þ a2nVn b2ð Þþ þ Vman1 tanh b an1V1 þ an2V2ð
þ þ annVn bnÞ u1
1
R1
ð10Þ
Also, if ‘E’ is the Energy Function, it must satisfy the
following condition [5].
oE
oV1
¼ KCp1
du1
dtð11Þ
where ‘K’ is a constant of proportionality and has the
dimensions of resistance.
Comparing (8) and (10) according to (11) yields
a11 ¼K
R11
; a21 ¼K
R21
; . . .; an1 ¼K
Rn1
ð12Þ
The values of the resistances of Fig. 1 can be easily
calculated by choosing a suitable value of ‘K’ and then
using (12). Analysis on similar lines can be performed to
obtain the values of the weights for the remaining neurons.
a11
V1+ a
12V
2 + ...+a
1nV
n
b1
Vi
Cpi
Rpi
R1i
Rni
x1
an1
V1+ a
n2V
2 + ...+a
nnV
n
bn x
n
Ni
Fig. 1 i-th neuron of the proposed feedback neural network circuit to
solve simultaneous linear equations in n-variables
434 Analog Integr Circ Sig Process (2011) 66:433–440
123
3 Energy function
3.1 Proof of the energy function
Differentiating the energy function of (9) w.r.t. Vi, we get
oE
oVi¼ Vma1i tanh b a11V1 þ a12V2 þ þ a1nVn b1ð Þ
þ Vma2i tanh b a21V1 þ a22V2 þ þ a2nVn b2ð Þ þ
þ Vmani tanh b an1V1 þ an2V2 þ þ annVn bnð Þ ui
Ri
ð13Þ
Comparing (8) and (10), we observed that
oE
oVi¼ KCpi
dui
dtð14Þ
where K = a1iR1i ; for all iFurther, the time derivative of the energy function is
given by
dE
dt¼XN
i¼1
oE
oVi
dVi
dt¼XN
i¼1
oE
oVi
dVi
dui
dui
dtð15Þ
Using (14) in (15) we get
dE
dt¼XN
i¼1
KCidui
dt
2dVi
duið16Þ
The transfer characteristics of the output opamp used in
Fig. 1 implements the activation function of the neuron.
With ui being the inverting terminal, it is monotonically
decreasing and it can be seen that [5],
dVi
dui 0 ð17Þ
Thereby resulting in
dE
dt 0 ð18Þ
With the equality being valid for
dui
dt¼ 0; for all i ð19Þ
Equation 18 shows that the energy function can never
increase with time which is one of the conditions for a valid
energy function. The second criterion viz. the energy
function must have a lower bound is also satisfied for the
circuit of Fig. 1 wherein it may be seen that V1, V2,…, Vn are
all bounded (as they are the outputs of opamps) amounting to
‘E’, as given in (9), having a defined lower bound.
3.2 Stable states of the network
Convergence of the network to the global minimum of the
Energy Function, which is exactly the solution of the set of
linear equations, and the fact that there are no other min-
ima, can be shown as follows.
For n variables, the second term in the energy function
expression (9) is significant only near the saturating values
of the opamp and is usually neglected [6]. The energy
function can therefore be expressed as
E ¼ Vm
Xn
i¼1
ln coshbXn
j¼1
aijVj bi
!ð20Þ
From which it follows that
oE
oV1
¼ Vma11 tanhb a11V1þa12V2þ þa1nVnb1ð Þ
þVma21 tanhb a21V1þa22V2þ þa2nVnb2ð Þþ þVman1 tanhb an1V1þan2V2þ þannVnbnð Þ
ð21Þ
oE
oV2
¼Vma12 tanhb a11V1þa12V2þþa1nVnb1ð Þ
þVma22 tanhb a21V1þa22V2þþa2nVnb2ð ÞþþVman2 tanhb an1V1þan2V2þþannVnbnð Þ
..
.
ð22Þ
oE
oVn¼Vma1n tanhb a11V1þa12V2þþa1nVnb1ð Þ
þVma2n tanhb a21V1þa22V2þþa2nVnb2ð Þþ þVmann tanhb an1V1þan2V2þþannVnbnð Þ
ð23Þ
For a stationary point, we have
oE
oV1
¼ 0
oE
oV2
¼ 0
..
.
oE
oVn¼ 0
9>>>>>>>>>>=>>>>>>>>>>;
ð24Þ
which yields,
a11 tanh b a11V1 þ a12V2 þ þ a1nVn b1ð Þþ a21 tanh b a21V1 þ a22V2 þ þ a2nVn b2ð Þ þ þ an1 tanh b an1V1 þ an2V2 þ þ annVn bnð Þ ¼ 0
ð25Þ
a12 tanh b a11V1 þ a12V2 þ þ a1nVn b1ð Þþ a22 tanh b a21V1 þ a22V2 þ þ a2nVn b2ð Þ þ þ an2 tanh b an1V1 þ an2V2 þ þ annVn bnð Þ ¼ 0
..
.
ð26Þ
Analog Integr Circ Sig Process (2011) 66:433–440 435
123
a1n tanh b a11V1 þ a12V2 þ þ a1nVn b1ð Þþ a2n tanh b a21V1 þ a22V2 þ þ a2nVn b2ð Þ þ þ ann tanh b an1V1 þ an2V2 þ þ annVn bnð Þ ¼ 0
ð27Þ
denoting
tanh b a11V1 þ a12V2 þ þ a1nVn b1ð Þ ¼ A1 ð28Þ
tanh b a21V1 þ a22V2 þ þ a2nVn b2ð Þ ¼ A2 ð29Þ
..
.
tanh b an1V1 þ an2V2 þ þ annVn bnð Þ ¼ An ð30Þ
Therefore, for a stationary point we have,
a11 a12 . . . a1n
a21 a22 . . . a2n
..
. ...
. . . ...
an1 an2 . . . ann
26664
37775
A1
A2
..
.
An
26664
37775 ¼
0
0
..
.
0
2664
3775 ð31Þ
This is a homogenous system of linear equations in
variables A1, A2,…, An. Since the coefficient matrix of
the set of Eq. 31 is the same as that of (1) which is
invertible, it follows that (31) will have a uniquely
determined solution which is the trivial solution of the
homogenous system.
Therefore,
A1
A2
..
.
An
26664
37775 ¼
0
0
..
.
0
2664
3775 ð32Þ
which results in,
a11V1 þ a12V2 þ þ a1nVn b1 ¼ 0
a21V1 þ a22V2 þ þ a2nVn b2 ¼ 0
..
.
an1V1 þ an2V2 þ þ annVn bn ¼ 0
9>>>>>=>>>>>;
ð33Þ
Thus, the energy function of the proposed neural
network has a unique stationary point which coincides
exactly with the solution of the given system of linear
equations.
This Energy Function can be visualized in 3-dimensions
for a 2—variable problem and is shown in Fig. 2. From the
plot, it can be seen that there exists only one minimum to
which the network must converge.
4 PSPICE simulation results
The proposed circuit was tested using PSPICE simulation
program for solving sets of 2, 3, 4, 5, 10 and 20 simulta-
neous linear equations. The application of the proposed
circuit to a chosen 3—variable problem (34) is presented
below.
2 1 1
3 2 1
1 1 2
24
35
V1
V2
V3
24
35 ¼
5
10
6
24
35 ð34Þ
The circuit to solve (34) as obtained from Fig. 1 is
presented in Fig. 3. As can be seen, some additional
circuitry is needed to generate the inputs to the non-linear
synapses. Routine analysis yields the following values of
the resistors at the input of the non-linear synapses.
Fig. 2 Typical energy function
plot for a 2—variable problem
436 Analog Integr Circ Sig Process (2011) 66:433–440
123
Re11 = 2.5 K, Re12 = Re13 = Re14 = 5 K, Re21 =
3.33 K, Re22 = 5 K, Re23 = 10 K, Re24 = 2.5 K,
Re31 = Re32 = 6 K, Re33 = Re34 = 3 K, R11 = R22 =
R33 =5 K, R12 = R13 = R23 = R31 = R32 = 10 K and
R21 = 3.33 K.
Algebraic analysis of (34) gives the solution as V1 =
-0.5 V, V2 = 5.5 V and V3 = 0.5 V. The results of
PSPICE simulation of the circuit of Fig. 3, shown in Fig. 5,
are found to match perfectly with the algebraic solution.
The initial node voltages were kept as V(1) = 10 mV,
V(2) = -1 mV and V(3) = 20 mV. For the purpose of
this simulation, the LM318 opamp from the Orcad library
in PSPICE was utilised. The transfer characteristics of the
opamp are presented in Fig. 4 below. The value of ‘b’ for
this opamp was measured to be 1.1 9 104 using SPICE
simulation.
The proposed circuit was further tested in PSPICE,
using other opamp models from Orcad library as well, for
solving systems of linear equations in 2, 4, 5, 10 and 20
variables. Results of simulation runs for these problems are
presented in Table 1. Each of the simulations was run using
various initial conditions in the millivolt range. As can be
seen from Table 1, the proposed network always converges
to the solution of the given system of linear equations.
5 Hardware implementation
Breadboad implementation of the proposed circuit was also
carried out. Apart from verification of the working of the
proposed circuit, the actual circuit realization also served
the purpose of testing the convergence of the circuit to the
solution starting from different initial conditions. The noise
present in any electronic circuit acts as a random initial
condition for the convergence of the neural circuit. Stan-
dard laboratory components viz. the lA741 opamp and
R11
R21
V1
a11V1+ a12V2 + a13V3
x1
u1
Rp1 Cp1
R22
R12
V2
b2
x2
u2
Rp2 Cp2
R33
R23
V3
b3
x3
u3
Rp3 Cp3
a21V1+ a22V2 + a23V3
a31V1+ a32V2 + a33V3
b1
R31
R32
R13
Re11
Re12
Re13
Re14
Re24
Re34
Re21
Re31
Re22
Re32
Re23
Re33
V1
V1
V1
V2
V2
V2
V3
V3
V3
N1
N2
N3
Fig. 3 The proposed circuit
applied to a 3—variable
problem
VIN
-1.0V -0.8V -0.6V -0.4V -0.2V -0.0V 0.2V 0.4V 0.6V 0.8V 1.0V
V(2)
-20V
-10V
0V
10V
20VFig. 4 Transfer characteristics
of the opamps used to realize
the neurons of Fig. 3
Analog Integr Circ Sig Process (2011) 66:433–440 437
123
resistances were used for the purpose. Circuits to solve sets
of equations in 2 and 3 variables were assembled and each
circuit was tested for 25 test runs. The results of hardware
implementation are presented in Table 2.
6 Discussion
This section deals with the monolithic implementation
issues of the proposed circuit. The PSPICE simulations
assumed that all operational amplifiers are identical, and
therefore, it is required to determine how deviations from
this assumption affect the performance of the network.
Effects of variations in component values from one neuron
to another were also investigated using Monte-Carlo
analysis in PSPICE. A 10% tolerance with Gaussian
deviation profile was put on the resistances used in the
circuit of Fig. 3. The analysis was carried out for 100 runs
and the Mean Deviation was found out to be 0.0497 and
Mean Sigma (Standard Deviation) was 0.1243. Offset
analysis was also carried out by incorporating random
offset voltages (in the range of 1–10 mV) to the opamps.
The Mean Deviation in this case was measured to be
-0.0512 and the Mean Sigma (Standard Deviation) was
0.172. As can be seen, the effects of mismatches and off-
sets on the overall precision of the final results are in an
acceptable range.
In fact, the drawbacks of the use of opamps in the
proposed circuit suggest that a real, large scale imple-
mentation for solving a system of equations in several
variables might be quite different. Alternative realizations
based on the differential Eq. 5 governing the system of
neurons are being investigated. An implementation based
on operational transconductance amplifiers (OTAs) would
appear to be one viable alternative; similarly, circuits using
MOS transistors operating in the subthreshold regime
would be another. In the latter case, the tanh(.) nonlinearity
is easily obtained because of the transistor’s current–volt-
age characteristic [7] and is expected to be the focus of
future work. However, the present manuscript focuses
mainly on the principle of such a network, and a VLSI
implementation is beyond the scope of this paper.
Further, scaling issues for the proposed circuit were also
investigated. For this purpose, two problems of 10 and 20
variables each were selected. These appear in Table 1 and
were solved using the network of Fig. 1. PSPICE simula-
tions was performed and the obtained simulation results
for the 10-variable problem were V1 = -1.9999, V2 =
1.0000, V3 = 3.0001, V4 = -1.0000, V5 = 1.9999, V6 =
7.0003, V7 = -3.0000, V8 = 3.9999, V9 = -5.0000 and
V10 = -4.0002 which agree very closely with the exact
mathematical solution.
7 Conclusions
In this paper we have described a novel approach to solve
n simultaneous linear equations in n variables, which uses
n neurons and n synapses. Each neuron requires one opamp
and each synapse is implemented using one comparator.
This results in significant reduction in hardware over the
existing schemes [1–4]. From ULSI implementation point
of view, more efficient designs are being explored in order
to reduce the circuitry required for each synapse.
Hardware implementations were carried out for 2 and 3
variable problems. The results were found to match with the
mathematical solution. The working of the proposed net-
work was also verified using PSPICE for various sample
problem sets of 2–20 simultaneous linear equations. While
the simulation results confirm the validity of the approach,
issues such as the network response to a system of equations
which do not have a unique solution are yet to be explored. It
may be mentioned that the technique can be extended to the
standard linear and quadratic programming problems.
Time
0s 5us 10us 15us 20us 25us 30us 35us 40us 45us 50us
V(1) V(2) V(3)
-2.0V
0V
2.0V
4.0V
6.0VFig. 5 Simulation results for
the chosen 3—variable problem
438 Analog Integr Circ Sig Process (2011) 66:433–440
123
Table 1 PSPICE simulation results for the proposed circuit applied to various problems
Analog Integr Circ Sig Process (2011) 66:433–440 439
123
Acknowledgements The authors are grateful to the anonymous
reviewers for their useful suggestions, which helped to improve the paper.
References
1. Jiang, D. (2004). Analog computing for real-time solution of time-
varying linear equations. In Proceedings of ICCCAS, Vol. 2,
pp. 27–29.
2. Xia, Y., Wang, J., & Hung, D. L. (1999). Recurrent neural
networks for solving linear inequalities and equations. IEEECircuits and Systems Society: I, 46(4), 452.
3. Cichocki, A., & Unbehauen, R. (1992). Neural networks for
solving systems of linear equations and related problems. IEEECircuits and Systems Society: I, 39(2), 124.
4. Wang, J. (1992). Electronic realization of recurrent neural network
for solving simultaneous linear equations. Electronics Letters,28(5), 493.
5. Rahman, S. A., Jayadeva, & Duttaroy, S. C. (1999). Neural
network approach to graph colouring. Electronics Letters, 35(14),
1173–1175.
6. Tank, D. W., & Hopfield, J. J. (1986). Simple neural optimization
networks: An A/D converter, signal decision network, and linear
programming circuit. IEEE Transactions on Circuits and SystemsCAS, 33(5), 533.
7. Newcomb, R. W., & Lohn, J. D. (1998) Analog VLSI for neural
networks. In The handbook of brain theory and neural networks.
Cambridge, MA: The MIT Press.
Syed Atiqur Rahman did
B.Sc.(Eng.) in Electrical Engi-
neering in 1988, and M.Sc.(Eng.)
in Electronics and Communica-
tion in 1994, from Aligarh
Muslim University, Aligarh. He
was appointed as Lecturer in
the Department of Electronics
Engineering at AMU, Aligarh in
1988. He joined as Reader in
Computer Engineering and then
Electronics Engineering in the
same University in 1997, and is
continuing at the latter post. His
fields of interest are Electronic
Circuits and Artificial Neural Networks.
Mohd. Samar Ansari received
B.Tech. degree in Electronics
Engineering from the Aligarh
Muslim University, Aligarh,
India, in 2001 and M.Tech.
degree in Electronics Engineer-
ing, with specialization in elec-
tronic circuit and system design,
in 2007. He is currently a
Lecturer in the Department of
Electronics Engineering, Ali-
garh Muslim University, where
he teaches electronic devices &
circuits and microelectronics.
His research interests include
micro-electronics, linear integrated circuit applications, neural
networks, and semiconductor physics.
Table 2 Hardware test results of the proposed circuit applied to 2 and 3 variable problems
Chosen set of
linear equations
Algebraic
solution
Hardware test results
(first five runs)
Standard deviation,
r (for 25 runs)
1 1
1 3
V1
V2
¼ 10
10
V1
V2
¼ 10
0
9:89
0:18
" #9:90
0:18
" #9:89
0:18
" #9:89
0:18
" #9:89
0:18
" #r1
r2
¼ 0:039
0:017
1 1 3
1 3 1
3 1 1
24
35
V1
V2
V3
24
35 ¼
10
10
10
24
35
V1
V2
V3
24
35 ¼
2
2
2
2435
2:04
1:98
2:00
264
375
2:01
1:97
2:00
264
375
2:04
1:98
2:01
264
375
2:04
1:98
2:00
264
375
2:04
1:97
2:01
264
375
r1
r2
r3
24
35 ¼
0:044
0:025
0:032
24
35
440 Analog Integr Circ Sig Process (2011) 66:433–440
123
Circuits Syst Signal Process (2011) 30:1029–1045DOI 10.1007/s00034-010-9261-x
DVCC-Based Non-linear Feedback Neural Circuitfor Solving System of Linear Equations
M. Samar Ansari · Syed Atiqur Rahman
Received: 26 October 2009 / Revised: 20 April 2010 / Published online: 14 January 2011© Springer Science+Business Media, LLC 2011
Abstract A neural circuit to solve a system of simultaneous linear equations is pre-sented. The circuit employs non-linear feedback to achieve a transcendental energyfunction that ensures fast convergence to the exact solution while enjoying reduc-tion in hardware complexity over existing schemes. A new building block for analogsignal processing, the digitally controlled differential voltage current conveyor (DC-DVCC) is introduced and is utilized for the non-linear synaptic interconnections be-tween neurons. The proof of the energy function has been given and it is shown thatthe gradient network converges exactly to the solution of the system of equations.PSPICE simulation results are presented for linear systems of equations of varioussizes and are found to be in close agreement with the algebraic solution. The use ofCMOS DC-DVCCs and operational amplifiers facilitates monolithic integration.
Keywords Neural network applications · Neural network hardware · Non-linearcircuits · Linear algebra · Linear equations · Digitally Controlled DVCC
1 Introduction
Solution of a system of simultaneous linear equations has been a primary goal forcomputation since the time of the abacus. References to word problems requiring thesolution of a system of linear equations appear in ancient Babylonian texts datingback to circa 300 BC [3]. Some examples of applications requiring the solution ofsystems of linear equations are curve fitting, electrical circuit analysis, multiple cor-
M.S. Ansari () · S.A. RahmanDepartment of Electronics Engineering, Aligarh Muslim University, Aligarh, Indiae-mail: [email protected]
S.A. Rahmane-mail: [email protected]
1030 Circuits Syst Signal Process (2011) 30:1029–1045
relation as well as real-time applications like real-time speech coding, image process-ing, stochastic modelling, and computer-aided realistic three-dimensional image syn-thesis [1, 3, 5, 10, 11]. Traditional methods for solving systems of linear equationstypically involve an iterative process [3], but long computational time limits their us-age. There is an alternative approach to the solution of this problem. It is to exploit theartificial neural networks (ANN’s), which can be considered as an analog computerrelying on a highly simplified model of neurons [1, 5, 10, 11]. Massively parallelprocessing and fast convergence, which are inherent to neural networks, are utilizedto reduce the time taken to arrive at the solution [1, 10, 11]. However, existing neuralnetworks have variable penalty parameters which decrease to zero as time increasesto infinity in order to get better accuracy of solution. Therefore, the time taken toarrive at the solution in such cases is not always short [1, 10].
Hardware solutions for solving a system of linear equations based on neural net-works have been put forward by Xia, Wang and Hung [11], Cichocki and Unbehauen[1] and Wang [10]. To solve an n-variable system of equations, the network of [10]uses three operational amplifiers, one capacitor and (n + 5) resistances to emulatea single neuron and the time to arrive at the solution is of the order of hundreds ofmilliseconds. The network of [1] provides a significantly improved solution time ofaround a microsecond but at the cost of increased hardware complexity. Each neuronin [1] comprises three weighted summers and an inverting integrator. The architec-ture of [11], which is a generalized neural network based implementation of Censorand Elfving’s method for linear inequalities, also uses an approach similar to [10],utilizing weighted adders and integrators to realize the neurons.
A new digital programmable architecture with fast convergence and reduced cir-cuit complexity is presented. The proposed architecture uses non-linear feedbackwhich leads to a new energy function that involves transcendental terms. This tran-scendental energy function is fundamentally different from the standard quadraticform associated with Hopfield network and its variations. The non-linear feedbackis realized by utilizing a new building block viz. the digitally controlled differen-tial voltage current conveyor (DC-DVCC). Along with presenting the analysis of theproposed circuit and the proof of the energy function, it is also shown that the stablestate of the network corresponds exactly with the solution of the given system of lin-ear equations. Further, the use of CMOS DC-DVCCs and opamps makes monolithicintegration viable.
This paper is organized as follows. A new digitally programmable analog build-ing block (DC-DVCC) is presented in Sect. 2. Section 3 contains the details of theproposed network along with the design equations. Proofs of the energy function andvalidity of the solution are given in Sect. 4. Section 5 presents the results of SPICEsimulation of the circuit applied to solve various sample equation sets. A discussionon VLSI implementability of the proposed circuit appears in Sect. 5. Some conclusiveremarks appear in Sect. 6.
2 Digitally Controlled DVCC
The differential voltage current conveyor (DVCC) was proposed in 1997 [2] as a fiveterminal device characterized by the following port relations (Fig. 1):
Circuits Syst Signal Process (2011) 30:1029–1045 1031
Fig. 1 Block representation ofthe DVCC
Fig. 2 CMOS implementation of DVCC
⎡⎢⎢⎢⎢⎢⎢⎣
VX
IY1
IY2
I+Z
I−Z
⎤⎥⎥⎥⎥⎥⎥⎦
=
⎡⎢⎢⎢⎢⎣
0 1 −1 0 00 0 0 0 00 0 0 0 0k 0 0 0 0k 0 0 0 0
⎤⎥⎥⎥⎥⎦
⎡⎢⎢⎢⎢⎢⎢⎣
IX
VY1
VY2
V +Z
V −Z
⎤⎥⎥⎥⎥⎥⎥⎦
(1)
While the voltage on the X-terminal follows the difference in voltages of terminalsY1 and Y2, a current injected at the X-terminal is replicated by a factor k to the Z-terminals. For the Z+ terminal, the direction of the conveyed current is the same asthat of the current flowing in the X-terminal whereas for the Z− terminal, the currentflows in the opposite direction. Ideally, k is unity.
One possible CMOS realization of the DVCC is shown in Fig. 2 [6]. As can beseen, the current at Z+ port will be the same as the current in the X terminal and thecurrent at the Z− terminal will have the same magnitude but opposite direction as theX port current.
Although both Z+ and Z− types of current outputs are mentioned in (1), theDVCCs used in the proposed network use only Z+ type of outputs. Therefore, wenext describe the proposed digitally controlled DVCC with only Z+ outputs. Thetechnique is to control the current transfer gain parameter k of the DVCC by replac-ing the Z terminal transistors of the DVCC with transistor arrays associated with
1032 Circuits Syst Signal Process (2011) 30:1029–1045
Fig. 3 CMOS realization of Digitally Controlled DVCC with gain k
switches [4]. The gain parameter k can take values from 1 to (2n − 1), where n rep-resents the number of transistor arrays. Actually, the transistor arrays implement acurrent summing network (CSN) at the Z terminal. The circuit of DC-DVCC ob-tained after suitable modifications in the circuit of Fig. 2 is presented in Fig. 3.
The CSN consists of n transistor pairs, whose NMOS and PMOS aspect ratios aregiven by:
NMOS:
(W
L
)
i
= 2i
(W
L
)
11, i = 0,1,2, . . . , (n − 1) (2)
PMOS:
(W
L
)
i
= 2i
(W
L
)
7, i = 0,1,2, . . . , (n − 1) (3)
Therefore, the current at the Z terminal, assumed flowing out of the DC-DVCC, canbe expressed by
IZ =n−1∑i=0
di2i (I7 − I11) (4)
Therefore, the proposed DC-DVCC provides a current transfer gain equal to
k = IZ
IX
=∑n−1
i=0 di2i (I7 − I11)
(I7 − I11)=
n−1∑i=0
di2i (5)
Parameter di represents the digital code-bit applied to the ith branch in the CSN.Depending upon its value, it enables or disables the current to flow in that particularbranch. It is instructive to note the numbering of the transistors in the CSN. Transis-tors labeled M8(i) and M12(i) refer to the PMOS and NMOS transistors in the CSNthat have been put there in the place of their counterparts of Fig. 2. Transistors MD8(i)
and MD12(i) are the actual digital control transistors as the digital control bits d0, d1and d2 are applied at their respective gate terminals.
Circuits Syst Signal Process (2011) 30:1029–1045 1033
For the DC-DVCC, comparator action can be achieved by putting RX → 0 i.e. bydirectly grounding the X-terminal [6]. For such a case, the current in the X-port willsaturate and can be written as
IX = Im tanhβ(VY1 − VY2) (6)
where β is the open-loop gain of the comparator (practically very high), ±Im are thesaturated output current levels of the comparator and V1,V2, . . . , Vn are the neuronoutputs. Equation (6) can also be written in an equivalent notation as
IX = gVm tanhβ(VY1 − VY2) (7)
where g is the transconductance from the input voltage ports (Y1 and Y2) to the outputcurrent port X and is governed by the resistance at the X port; and ±Vm are thebiasing voltages of the DVCC-based comparator.
By virtue of DVCC action, this current will be transferred to the Z+ ports as
I+Z = kIX = kgVm tanhβ(VY1 − VY2) (8)
Therefore, the current at the Z terminal can be made a digitally controlled scaledreplica of the X port current. This property will be utilized in the design of the multi-output DVCCs needed for the proposed circuit. It may be mentioned that multiple Z+outputs can be obtained by repeating the output stage comprising transistors M8(i)
and M12(i).
3 Proposed Neural Network
Let the simultaneous linear equations to be solved are
AV = B (9)
where
A =
⎡⎢⎢⎣
a11 a12 . . . a1n
a21 a22 . . . a2n...
... . . ....
an1 an2 . . . ann
⎤⎥⎥⎦ (10)
B =
⎡⎢⎢⎣
b1b2...
bn
⎤⎥⎥⎦ (11)
V =
⎡⎢⎢⎣
V1V2...
Vn
⎤⎥⎥⎦ (12)
1034 Circuits Syst Signal Process (2011) 30:1029–1045
Fig. 4 First neuron of the proposed digitally programmable, feedback neural circuit to solve simultaneouslinear equations in n-variables
where V1,V2, . . . , Vn are the variables and aij and bi are constants. We will assumethat the coefficient matrix A is invertible, and hence, the system of linear equations(9) is consistent and not under-determined. In other words, the linear system (9) hasa uniquely determined solution.
The proposed neural network based circuit to solve the system of equations of (9)is presented in Fig. 4. As can be seen from Fig. 4, individual equations from the setof equations to be solved are passed through non-linear synapses which are realizedusing multi-output DC-DVCC based comparators.
The outputs of the comparators are fed to neurons having weighted inputs. Theseweighted neurons are realized by using opamps where the scaled currents comingfrom various comparators act as weights. Rpi and Cpi are the input resistance andcapacitance of the opamp corresponding to the ith neuron. These parasitic compo-nents are included to model the dynamic nature of the opamp.
Node equation for node ‘Ni ’ gives the equation of motion of the ith neuron as
Cpi
dui
dt= k1iIX1 + k2iIX2 + · · · + kniIXn − ui
[1
Rpi
](13)
where ui is the internal state of the ith neuron and kji is the current scaling factor atthe j th output of ith DVCC. Using (7) in (13) results in
Cpi
dui
dt= k1igVm tanhβ(a11V1 + a12V2 + · · · + a1nVn − b1)
+ k2igVm tanhβ(a21V1 + a22V2 + · · · + a2nVn − b2) + · · ·+ knigVm tanhβ(an1V1 + an2V2 + · · · + annVn − bn) − ui
Rpi
(14)
Circuits Syst Signal Process (2011) 30:1029–1045 1035
Moreover, as has been shown in Sect. 4, the network in Fig. 4 can be associated withan Energy Function ‘E’ given by
E = Vm
n∑i=1
ln coshβ
(n∑
j=1
aijVj − bi
)−
n∑i=1
1
Ri
∫ Vi
0ui dVi (15)
From (15), it follows that
∂E
∂V1= Vma11 tanhβ(a11V1 + a12V2 + · · · + a1nVn − b1)
+ Vma21 tanhβ(a21V1 + a22V2 + · · · + a2nVn − b2) + · · ·
+ Vman1 tanhβ(an1V1 + an2V2 + · · · + annVn − bn) − u1
[1
Rp1
](16)
Also, if ‘E’ is the energy function, it must satisfy the following condition [8].
∂E
∂Vi
= KRCpi
dui
dt(17)
where KR is a constant of proportionality having the dimensions of resistance and isnormalized to (1/g) for simplicity.
Comparing (14) and (16) according to (17) yields
⎡⎢⎢⎣
k11 k12 . . . k1n
k21 k22 . . . k2n...
... . . ....
kn1 kn2 . . . knn
⎤⎥⎥⎦ =
⎡⎢⎢⎣
a11 a12 . . . a1n
a21 a22 . . . a2n...
... . . ....
an1 an2 . . . ann
⎤⎥⎥⎦ (18)
4 Energy Function
4.1 Proof of the Energy Function
Differentiating the energy function of (15) w.r.t. Vi , we get
∂E
∂Vi
= Vma1i tanhβ(a11V1 + a12V2 + · · · + a1nVn − b1)
+ Vma2i tanhβ(a21V1 + a22V2 + · · · + a2nVn − b2) + · · ·+ Vmani tanhβ(an1V1 + an2V2 + · · · + annVn − bn) − ui
Rpi
(19)
Further, the time derivative of the energy function is given by
dE
dt=
N∑i=1
∂E
∂Vi
dVi
dt=
N∑i=1
∂E
∂Vi
dVi
dui
dui
dt(20)
1036 Circuits Syst Signal Process (2011) 30:1029–1045
Using (12) in (15) we get
dE
dt=
N∑i=1
KCi
(dui
dt
)2dVi
dui
(21)
The transfer characteristics of the output opamp used in Fig. 4 implements the activa-tion function of the neuron. With ui being the inverting terminal, it is monotonicallydecreasing and it can be seen that [8]
dVi
dui
≤ 0 (22)
thereby resulting in
dE
dt≤ 0 (23)
with the equality being valid for
dui
dt= 0; for all i (24)
Equation (23) shows that the energy function can never increase with time which isone of the conditions for a valid energy function. The second criterion viz. the energyfunction must have a lower bound is also satisfied for the circuit of Fig. 4 wherein itmay be seen that V1,V2, . . . , Vn are all bounded (as they are the outputs of opamps)amounting to ‘E’, as given in (15), having a defined lower bound.
4.2 Stable States of the Network
Convergence of the network to the global minimum of the energy function, which isexactly the solution of the set of linear equations, and the fact that there are no otherminima, can be shown as follows.
For n variables, the second term in the energy function expression (15) is signif-icant only near the saturating values of the opamp and is usually neglected [9]. Theenergy function can therefore be expressed as
E = Vm
n∑i=1
ln coshβ
(n∑
j=1
aijVj − bi
)(25)
from which it follows that
∂E
∂V1= Vma11 tanhβ(a11V1 + a12V2 + · · · + a1nVn − b1)
+ Vma21 tanhβ(a21V1 + a22V2 + · · · + a2nVn − b2) + · · ·+ Vman1 tanhβ(an1V1 + an2V2 + · · · + annVn − bn) (26)
Circuits Syst Signal Process (2011) 30:1029–1045 1037
∂E
∂V2= Vma12 tanhβ(a11V1 + a12V2 + · · · + a1nVn − b1)
+ Vma22 tanhβ(a21V1 + a22V2 + · · · + a2nVn − b2) + · · ·+ Vman2 tanhβ(an1V1 + an2V2 + · · · + annVn − bn) (27)
...
∂E
∂Vn
= Vma1n tanhβ(a11V1 + a12V2 + · · · + a1nVn − b1)
+ Vma2n tanhβ(a21V1 + a22V2 + · · · + a2nVn − b2) + · · ·+ Vmann tanhβ(an1V1 + an2V2 + · · · + annVn − bn) (28)
For a stationary point, we have
∂E
∂V1= 0
∂E
∂V2= 0
...
∂E
∂Vn
= 0
⎫⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎬⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎭
(29)
which yields
a11 tanhβ(a11V1 + a12V2 + · · · + a1nVn − b1)
+ a21 tanhβ(a21V1 + a22V2 + · · · + a2nVn − b2) + · · ·+ an1 tanhβ(an1V1 + an2V2 + · · · + annVn − bn) = 0 (30)
a12 tanhβ(a11V1 + a12V2 + · · · + a1nVn − b1)
+ a22 tanhβ(a21V1 + a22V2 + · · · + a2nVn − b2) + · · ·+ an2 tanhβ(an1V1 + an2V2 + · · · + annVn − bn) = 0 (31)
...
a1n tanhβ(a11V1 + a12V2 + · · · + a1nVn − b1)
+ a2n tanhβ(a21V1 + a22V2 + · · · + a2nVn − b2) + · · ·+ ann tanhβ(an1V1 + an2V2 + · · · + annVn − bn) = 0 (32)
Denoting
tanhβ(a11V1 + a12V2 + · · · + a1nVn − b1) = A1 (33)
tanhβ(a21V1 + a22V2 + · · · + a2nVn − b2) = A2 (34)...
tanhβ(an1V1 + an2V2 + · · · + annVn − bn) = An (35)
1038 Circuits Syst Signal Process (2011) 30:1029–1045
Therefore, for a stationary point we have⎡⎢⎢⎣
a11 a12 . . . a1n
a21 a22 . . . a2n...
... . . ....
an1 an2 . . . ann
⎤⎥⎥⎦
⎡⎢⎢⎣
A1A2...
An
⎤⎥⎥⎦ =
⎡⎢⎢⎣
00...0
⎤⎥⎥⎦ (36)
This is a homogeneous system of linear equations in variables A1,A2, . . . ,An. Sincethe coefficient matrix of the set of (36) is the same as that of (9) which is invertible, itfollows that (36) will have a uniquely determined solution which is the trivial solutionof the homogeneous system.
Therefore, ⎡⎢⎢⎣
A1A2...
An
⎤⎥⎥⎦ =
⎡⎢⎢⎣
00...0
⎤⎥⎥⎦ (37)
which results ina11V1 + a12V2 + · · · + a1nVn − b1 = 0
a21V1 + a22V2 + · · · + a2nVn − b2 = 0...
an1V1 + an2V2 + · · · + annVn − bn = 0
⎫⎪⎪⎪⎬⎪⎪⎪⎭
(38)
Thus, the energy function of the proposed neural network has a unique stationarypoint which coincides exactly with the solution of the given system of linear equa-tions.
This energy function can be visualized in three dimensions for a two-variableproblem and is shown in Fig. 5. From the plot, it can be seen that there exists onlyone minimum, to which the network must converge.
Fig. 5 Typical energy function plot for a two-variable problem
Circuits Syst Signal Process (2011) 30:1029–1045 1039
5 PSPICE Simulation Results
The proposed circuit was tested using PSPICE simulation program for solving setsof two, three, four, five and 10 simultaneous linear equations. The application of theproposed circuit to a chosen three-variable problem (39) is presented below.
⎡⎣
2 1 13 2 11 1 2
⎤⎦
⎡⎣
V1V2V3
⎤⎦ =
⎡⎣
5106
⎤⎦ (39)
The circuit to solve (39) as obtained from Fig. 4 is presented in Fig. 6. The values ofthe current scaling coefficients kji are given as
⎡⎣
k11 k12 k13k21 k22 k23k31 k32 k33
⎤⎦ =
⎡⎣
2 1 13 2 11 1 2
⎤⎦ (40)
For implementing these values of kji appropriate values of the control words wereselected. For the first DC-DVCC, the three control words were kept as [0 1 0], [0 0 1]
Fig. 6 The proposed circuit applied to a three-variable problem
1040 Circuits Syst Signal Process (2011) 30:1029–1045
Fig. 7 Simulation results for the chosen three-variable problem
and [0 0 1]. Similarly, the control words for the second DC-DVCC were [0 1 1],[0 1 0] and [0 0 1]. Finally, for the third DC-DVCC the control words were [0 0 1],[0 0 1] and [0 1 0]. Thus, a proper selection of the control words can be used toachieve any current scaling factor between 1 and 7 and therefore this particular re-alization is restricted to solving systems of equations with coefficients in that range.However, that range may be widened further by having four control bits in the CSN.
As can be seen, some additional circuitry is needed to generate the inputs to thenon-linear synapses. Routine analysis yields the following values of the resistors atthe input of the non-linear synapses.
Re11 = 2.5 K, Re12 = Re13 = Re14 = 5 K, Re21 = 3.33 K, Re22 = 5 K,
Re23 = 10 K, Re24 = 2.5 K, Re31 = Re32 = 6 K, Re33 = Re34 = 3 K,
Rc11 = Rc21 = Rc31 = 1 K, Rc12 = 4 K, Rc22 = 9 K, Rc32 = 5 K.
Algebraic analysis of (39) gives the solution as V1 = −0.5 V, V2 = 5.5 V andV3 = 0.5 V. The results of PSPICE simulation of the circuit of Fig. 6, shown inFig. 7, are found to match perfectly with the algebraic solution. The initial nodevoltages were kept as V (1) = 10 mV, V (2) = −10 mV and V (3) = 20 mV. The cir-cuit for DVCC was taken from [6] and standard 0.5 micron CMOS parameters wereused for simulation purposes. The DVCCs were biased with ±2.5 V supplies. Forthe opamp, use was made of the LMC7101A CMOS opamp from National Semicon-ductor. The sub-circuit file for this opamp is available in Orcad Model Library. The
Circuits Syst Signal Process (2011) 30:1029–1045 1041
Fig. 8 Transfer characteristics of the DC-DVCC used to realize the comparator
biasing voltages for the opamps were taken to be ±15 V. The transfer characteris-tics of the DVCC-based comparator, shown in Fig. 8, were also plotted using SPICEsimulation and value of ‘β’ was found to be 2.4 × 103.
The proposed circuit was further tested in PSPICE for solving systems of linearequations in two, four, five, and ten variables. Results of simulation runs for theseproblems are presented in Table 1. Each of the simulations was run using variousinitial conditions in the millivolt range. As can be seen from Table 1, the proposednetwork always converges to the solution of the given system of linear equations.As can be seen, the obtained results are quite near the algebraic solutions with themaximum error being approximately 6% and the average error being 0.49%.
6 Performance Evaluation
This section deals with the effect of component mismatches and device non-idealitieson the accuracy of the solution obtained from the circuit proposed.
If the proposed circuit is implemented using discrete resistances, variations in thevalues of the resistances need to be considered keeping in mind the tolerances asso-ciated with discrete resistors. However, if the proposed network is targeted for mono-lithic integration, random variations in the resistance values need not be considered.In that case, all the resistances are expected to deviate from their assigned values bythe same factor [7]. Therefore, a worthy performance appraisal of the circuit couldbe obtained by testing the circuit with all resistances having the same percentage de-viation from their assigned values. Such an assessment of the quality of the obtainedsolution is presented in Table 2 from which it can be seen that the percentage error
1042 Circuits Syst Signal Process (2011) 30:1029–1045
Table 1 PSPICE Simulation results for the proposed circuit applied to different systems of linear equa-tions
[A] [B] AlgebraicSolution
Simulated Results(Using PSPICE)
Percentage Errorin the solution (%)
[V] [V]
[1 22 1
] [3.54
] [1.51
] [1.511.02
] [ −0.66−2.00
]
[2 1 13 2 11 1 1
] [5106
] [−0.55.50.5
] [−0.525.540.48
] [−4.00−0.724.00
]
⎡⎣
2 3 6 13 2 5 42 5 5 22 4 2 5
⎤⎦
⎡⎣
48.7567.554
59.25
⎤⎦
⎡⎣
3.51.55
7.25
⎤⎦
⎡⎣
3.511.535.097.19
⎤⎦
⎡⎣
−0.28−2.00−1.800.83
⎤⎦
⎡⎢⎢⎣
2 3 9 2 52 6 9 9 52 6 2 4 52 4 7 8 35 3 6 3 5
⎤⎥⎥⎦
⎡⎢⎢⎣
−54.9−91.0−54.9−74.8−70.0
⎤⎥⎥⎦
⎡⎢⎢⎣
−6.7−4.3−2.8−3.30.64
⎤⎥⎥⎦
⎡⎢⎢⎣
−6.65−4.33−2.63−3.190.63
⎤⎥⎥⎦
⎡⎢⎢⎣
0.75−0.696.073.331.56
⎤⎥⎥⎦
⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣
1 2 1 3 4 2 1 1 1 22 1 1 3 1 2 2 1 2 31 1 4 3 3 1 1 4 4 14 2 1 5 3 3 1 1 2 21 1 5 1 2 1 2 2 5 23 3 1 2 1 1 5 2 1 15 5 1 4 1 1 3 4 2 21 1 1 2 2 2 3 3 4 14 2 2 1 3 2 5 4 3 21 2 3 1 2 3 1 2 3 4
⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦
⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣
10−11102
−7−9−8−3−37
⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦
⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣
−213
−127
−34
−5−4
⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦
⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣
−1.911.043.05
−0.951.947.04
−3.063.99
−4.91−3.95
⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦
⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣
4.50−4.00−1.665.003.00
−0.57−2.000.251.801.25
⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦
Table 2 Effect of variation inresistances on the obtainedresults
Percentage Variationin Resistances
Simulated Results(Using PSPICE)[V]
Percentage Error inthe Solution (%)
+2%
[−0.5085.6020.499
] [ +1.6+1.85−0.2
]
+5%
[−0.5095.6190.501
] [ +1.8+2.16+0.2
]
+10%
[−0.5145.6810.507
] [ +2.8+3.29+1.4
]
−2%
[−0.5075.6010.505
] [ +1.4+1.83+1
]
−5%
[−0.5185.6790.506
] [ +3.6+3.25+1.2
]
−10%
[−0.5165.6780.504
] [ +3.2+3.23+0.8
]
Circuits Syst Signal Process (2011) 30:1029–1045 1043
Table 3 Effect of gains of theDVCC-based comparators onthe solution quality
X-terminal resistancesof the DVCCs ()
Simulated Results(Using PSPICE)[V]
Percentage Errorin the solution (%)
5
[−0.5125.6100.500
] [2.42.00.0
]
10
[−0.5145.6170.510
] [2.82.1
2.00
]
20
[−0.5205.6200.511
] [−4.002.182.2
]
in the solution varies with the percentage errors specified in the values of the resis-tances and even for ±10% variation in the values of the resistors, the solution pointchanges by approximately 3%.
Further, the gains of the DVCC-based comparators were varied to investigate theireffect on the solution quality. Toward that end, a resistance was connected betweenthe X-terminal of each DVCC and ground. Ideally, these resistances were assumed tohave zero value. However, by assigning different values to these resistances, the gainsof the DVCCs were varied. The solutions, as obtained for the three-variable problemof Fig. 6, are presented in Table 3. It can be seen that small variations in the gains ofthe comparators do not affect the quality of the solution.
Next, the effect of offset voltages in the DVCC-based comparators was explored.Offset voltages were applied at the Y2 inputs of the DVCCs of Fig. 6 and the results ofPSPICE simulations were compared with the algebraic solution as given in Table 4.As can be seen, the offset voltages of the comparators do not affect the obtainedsolutions to any appreciable extent. However, the error does tend to increase withincreasing offset voltages.
Finally, offset voltages for the opamps were also considered. Offset voltages wereapplied at the non-inverting inputs of the opamps of Fig. 6 and the results of PSPICEsimulations were compared with the algebraic solution as given in Table 5. As can beseen, the offset voltages of the opamps have little effect on the obtained solutions.
7 Conclusion
In this paper we have described a novel approach to solve n simultaneous linearequations in n variables, which uses n neurons and n synapses. Each neuron requiresone opamp and each synapse is implemented using one comparator. The compara-tors were realized using a new digitally controlled DVCC. This results in significantreduction in hardware over the existing schemes [1, 5, 10, 11]. From VLSI imple-mentation point of view, use of CMOS DC-DVCCs and opamps facilitates viabilityfor monolithic integration.
The working of the proposed network was verified using PSPICE for various sam-ple problem sets of two to 10 simultaneous linear equations. While the simulation
1044 Circuits Syst Signal Process (2011) 30:1029–1045
Table 4 Effect of offsetvoltages of the DVCC-basedcomparators on the solutionquality
Offset Voltage appliedat Y2 input of the DVCCs(mV)
Simulated Results(Using PSPICE)[V]
Percentage Errorin the Solution (%)
−5
[−0.4985.6150.497
] [ −0.42.09−0.6
]
−10
[−0.5055.6200.506
] [1.02.181.2
]
−15
[−0.5125.6310.511
] [2.42.382.2
]
+5
[−0.4965.6110.494
] [ −0.82.01−1.2
]
+10
[−0.5035.6270.508
] [0.62.301.6
]
+15
[−0.5195.6290.510
] [3.82.34
2
]
Table 5 Effect of offsetvoltages of opamps on solutionquality
Offset Voltage appliedat the non-inverting inputof opamps (mV)
Simulated Results(Using PSPICE)[V]
Percentage Error inthe Solution (%)
−5
[−0.4945.6170.507
] [−1.22.121.4
]
−10
[−0.5045.5810.510
] [0.81.47
2
]
−15
[−0.5145.5430.491
] [2.80.78−1.8
]
+5
[−0.4995.6420.497
] [ −0.22.58−0.6
]
+10
[−0.5065.6680.505
] [1.23.051.0
]
+15
[−0.5125.6930.507
] [2.43.511.4
]
results confirm the validity of the approach, issues such as the network response toa system of equations which do not have a unique solution are yet to be explored.It may be mentioned that the technique can be extended to the standard linear andquadratic programming problems.
Circuits Syst Signal Process (2011) 30:1029–1045 1045
References
1. A. Cichocki, R. Unbehauen, Neural networks for solving systems of linear equations and relatedproblems. IEEE Trans. Circuits Syst.—I 39(2), 124–138 (1992)
2. H.O. Elwan, A.M. Soliman, Novel CMOS differential voltage current conveyor and its applications.IEE Proc. Circuits Devices Syst. 144(3), 195–200 (1997)
3. J.L. Gustafson, The quest for linear equation solvers and the invention of electronic digital computing,in Proc. JVA’06, IEEE John Vincent Atanasoff 2006 International Symposium on Modern Computing(2006)
4. T.M. Hassan, S.A. Mahmoud, Fully programmable universal filter with independent gain-ωo-Q con-trol based on new digitally programmable CMOS CCII. J. Circuits Syst. Comput. 18(5), 875–897(2009)
5. D. Jiang, Analog computing for real-time solution of time-varying linear equations. Proc. ICCCAS 2,27–29 (2004)
6. S. Maheshwari, A canonical voltage-controlled VM-APS with a grounded capacitor. Circuits Syst.Signal Process. 27, 123–132 (2008)
7. J.D. Plummer, M.D. Deal, P.B. Griffin, Silicon VLSI Technology: Fundamentals, Practices and Mod-eling (Pearson Education, Upper Saddle River, 2009)
8. S.A. Rahman, Jayadeva, S.C. Duttaroy, Neural network approach to graph colouring. Electron. Lett.35(14), 1173–1175 (1999)
9. D.W. Tank, J.J. Hopfield, Simple neural optimization networks: an A/D converter, signal decisionnetwork, and linear programming circuit. IEEE Trans. Circuits Syst. CAS 33(5), 533–541 (1986)
10. J. Wang, Electronic realization of recurrent neural network for solving simultaneous linear equations.Electron. Lett. 28(5), 493–495 (1992)
11. Y. Xia, J. Wang, D.L. Hung, Recurrent neural networks for solving linear inequalities and equations.IEEE Trans. Circuits Syst.—I 46(4) (1999)
International Journal of Computer Applications (0975 – 8887)
Volume 39– No.2, February 2012
44
Non-linear Feedback Neural Network for Solution of
Quadratic Programming Problems
Mohd. Samar Ansari
Department of Electronics Engineering Aligarh Muslim University
Aligarh, India
Syed Atiqur Rahman Department of Electronics Engineering
Aligarh Muslim University Aligarh, India
ABSTRACT
This paper presents a recurrent neural circuit for solving
quadratic programming problems. The objective is
tominimize a quadratic cost function subject to
linearconstraints. The proposed circuit employs non-
linearfeedback, in the form of unipolar comparators, to
introducetranscendental terms in the energy function ensuring
fastconvergence to the solution. The proof of validity of the
energy function is also provided. The hardware complexity of
the proposed circuit comparesfavorably with other proposed
circuits for the same task. PSPICE simulation results
arepresented for a chosen optimization problem and are
foundto agree with the algebraic solution.
General Terms
Neural Networks, Quadratic Programming Problem.
Keywords
Dynamical Systems, Non-Linear Synapse, Feedback
Networks.
1. INTRODUCTION Quadratic programming problem (QPP) is the problem of
optimizing (minimizing or maximizing) a quadratic function
of several variables subject to linear constraints on these
variables. Such problems arise naturally in a variety of
applications, such as structural analysis [1], optimal control
[2], plastic analysis [3], antenna array pattern synthesis [4],
geometric optimization [5], propulsion physics [6], multi-
commodity networks [7], etc. Moreover, in the discipline of
constrained optimization, problems with nonlinear objective
functions are usually approximated by a second-order system
and solved by a standard quadratic programming technique.
Traditional methods for solving quadratic programming
problems typically involve an iterative process, but long
computational time limits their usage. An alternative approach
to solution of this problem is to exploit the artificial neural
networks (ANN's) which can be considered as an analog
computer relying on a highly simplified model of neurons [8].
ANN's have been applied to several classes of constrained
optimization problems and have shown promise for solving
such problems more effectively. For example, the Hopfield
neural network has proven to be a powerful tool for solving
some of the optimization problems. Tank and Hopfield first
proposed a neural network for solving mathematical
programming problems, where a linear programming problem
(LPP) was mapped into a closed-loop network [9]. Later, the
dynamical approach was extended for solving quadratic
programming problems. Over the past two decades several
neural-network architectures for solving quadratic
programming problems have been proposed by Kennedy &
Chua [10], Maa&Shanblatt[11], Chen & Fang [12], Wu et
al.[13] and Xia [14]. More recently, Malek&Alipour proposed
a recurrent neural network that is able to solve quadratic
programming problems [15] without needing to set network
parameters thereby reducing the number of analog multipliers
required.
In this paper, a hardware solution to the problem of solving a
quadratic programming problem is presented. The proposed
architecture uses non-linear feedback which leads to a new
energy function that involves transcendental terms. This
transcendental energy function is fundamentally different
from the standard quadratic form associated with Hopfield
network and its variants. To solve a QPP in n variables with m
constraints, the circuit requires n opamps,m unipolar
comparators and (n2+mn) resistances thereby causing the
hardware complexity of the proposed network to compare
favorably with the existing hardware implementations. It may
be mentioned that a similar approach of using non-linear
synaptic interconnections between neurons has also been
employed to solve systems of simultaneous linear equations
[16] and linear programming problems [17].
The remainder of this paper is arranged as follows. A brief
review of relevant technical literature on the solution of QPP
using neural network based methods is presented in Section-2.
Section-3 outlines the mathematical formulation of the basic
problem and details of the proposed network. Section-4
contains explanation of the energy function and the proof of
its validity. Section-5 contains the circuit implementation of
the proposed network for a set of sample problem in two
variables. PSPICE simulation results of the proposed circuit
are also presented. Issues that are expected to arise in actual
monolithic implementations are discussed in Section-6.
Concluding remarks are presented in Section-7.
2. EXISTING NEURAL NETWORKS
FOR QPP Various methods to solve QPP by employing neural network
approaches are available in the technical literature. Kennedy
& Chua extended the Tank and Hopfield network by
developing a neural network for solving nonlinear
programming problems, by satisfaction of the Karush–Kuhn–
Tucker optimality conditions [10]. However, the need to set a
penalty parameter means that the network can generate
approximate solutions only and implementation problems
arise when the penalty parameter is large. Each variable
amplifier comprises of 2 opamps, 2 resistors and 1 capacitor
whereas for satisfying each constraint, the constraint amplifier
employs 3 opamps, 2 resistors and 1 diode [10]. Wang
proposed a recurrent neural network for solving QPPs with
equality constraints. The network is asymptoticallystable and
is able to generate optimal solutions to quadratic programs
with equality constraints. An opamp based circuit realization
International Journal of Computer Applications (0975 – 8887)
Volume 39– No.2, February 2012
45
of the network is also presented which requires (n+m) neurons
for solving a QPP in n variables with m constraints. Each
neuron is made up of a summer, an integrator, and an inverter
consuming 3 opamps, 1 capacitor and (n+5) resistors [18].
Wang's network is not suitable for real time applications as it
takes around 50 ms to arrive at the solution [18]. A rigorous
analysis of the prominent neural networks for QPP, available
till that time (1992), is presented in [11]. Forti&Tesi presented
new conditions capable of ensuring existence, uniqueness, and
global asymptotic stability of the equilibrium point for
Kennedy and Chua's network [19]. Wu et al. proposed two
neural network models for solving LPP and QPP, the
convergence of which was not dependent on the network
parameters [13].
Around the same time, Xia also put forward a neural network
capable of solving both LPP and QPP in which no parameter
tuning was necessary. Moreover, the actual hardware
implementation was somewhat simplified, as compared to its
contemporaries, because of the fact that no analog multipliers
were required for the variables [14]. To solve a QPP in n
variables with m constraints, Xia's network consisted of
(2m2+4mn) amplifiers, (2m2+4mn+3m+3) summers, (n+m)
integrators, and n limiters. Tao, Cao and Sun further
simplified the network of Xia [14], and reduced the system
complexity [20]. More recently, Liu and Wang presented a
one layer feedback neural network with a discontinuous hard-
limiting activation function for solving QPP in which the
number of neurons is the same as the number of decision
variables [21]. Each neuron in [21] is composed of two
adders, (3n+1) resistors, one limiter and an integrator.
Although significant reduction in circuit complexity is
achieved, the time that the circuit takes to arrive at the correct
solution is of the order of seconds thereby making the circuit
unsuitable for applications requiring fast solution times. A
comprehensive bibliography of the technical literature related
to QPP can be found in [22].
3. PROPOSED CIRCUIT Let the second-orderfunction to be minimized be
𝐹 =
𝑉1
𝑉2
⋮𝑉𝑛
𝑇
c11 c12 … c1n
c21 c22 … c2n
⋮ ⋮ … ⋮cn1 cn2 … cnn
𝑉1
𝑉2
⋮𝑉𝑛
(1)
subject tothe following linear constraints
a11 a12 … a1n
a21 a22 … a2n
⋮ ⋮ … ⋮am1 am2 … amn
𝑉1
𝑉2
⋮𝑉𝑛
≤
b1
b2
⋮bm
(2)
whereV1, V2,…, Vn are the variables, and aij, cij and bi (i = 1, 2,
…, m; j = 1, 2, …, n) are constants. The proposed
neuralnetwork based circuit to minimize the quadratic
function given in (1) in accordance with the constraints of (2)
is presented in Fig. 1. As can be seen from Fig. 1, individual
equations from the set of equations to be solved are passed
through non-linear synapses which are realized using unipolar
comparators comprising of operational amplifiers and diodes.
Rp1 and Cp1are the input resistance and capacitance of the
opamp that is used to emulate the functionality of a neuron.
These parasitic components are included to model the
dynamic nature of the opamp. The outputs of the comparators
are fed to neurons having weighted inputs. The neurons are
realized by using opamps and the weights are implemented
using resistances. The currents arriving to the neuron from
various synapses get added up at the input of the neuron.
Fig 1: First neuron of the proposed feedback neural
network circuit to solve a quadratic programming
problem in n variables with m linear constraints.
Graphical representation of the transfer characteristics for a
bipolar comparator is shown in Fig 2(a) from where it can be
seen that the comparator output saturates at ±Vm when the two
inputs differ by more than few millivolts in magnitude.
Unipolar transfer characteristics can be obtained using an
opamp (the transfer characteristics of which can be modeled
by (3) by employing a diode as depicted in Fig. 3, the diode
essentially `trimming' one half of the transfer characteristic
curve, which are shown in Fig. 2(b) and can be
mathematically modeled by (4). As is explained in the next
section, such unipolar comparator characteristics are utilized
to obtain an energy function which acts to bring the neuronal
states to the feasible region.
Fig 2: Transfer characteristics of (a) bipolar; and
(b)unipolar; comparators
Fig 3: Obtaining unipolar comparator characteristics
using an opamp and a diode
𝑥 = 𝑉𝑚 𝑡𝑎𝑛ℎ 𝛽 𝑉𝑖 − 𝑉𝑗 (3)
𝑥 =1
2𝑉𝑚 𝑡𝑎𝑛ℎ 𝛽 𝑉𝑖 − 𝑉𝑗 + 1 (4)
Using (4), the output of the i-th unipolar comparator in Fig. 1
can be given by (5) where βis the open-loop gain of the
comparator (practically very high), ±Vm are the output voltage
levels of the comparator and V1, V2,…,Vn are the neuron
outputs.
𝑥𝑖 =𝑉𝑚
2 𝑡𝑎𝑛ℎ 𝛽 𝑎𝑖1𝑉1 + 𝑎𝑖2𝑉2 + ⋯ + 𝑎𝑖𝑛𝑉𝑛 − 𝑏𝑖 + 1 (5)
International Journal of Computer Applications (0975 – 8887)
Volume 39– No.2, February 2012
47
Using (15) in (19) we get
𝑑𝐸
𝑑𝑡= 𝐾𝐶𝑖
𝑁
𝑖=1
𝑑𝑢𝑖
𝑑𝑡
2 𝑑𝑉𝑖
𝑑𝑢𝑖
20
The transfer characteristics of the output opamp used to
implement the neurons in Fig. 1 implements the activation
function of the neuron and can be written as
𝑉𝑖 = 𝑓 𝑢𝑖 21
Fig 5: Transfer characteristics of the opamp used to
realize the neurons
whereVi denotes the output of the opamp and ui corresponds
to the internal state at the inverting terminal. The function f is
typically a saturating, monotonically decreasing one, as
shown in Fig. 5, and therefore [16],
𝑑𝑉𝑖
𝑑𝑢𝑖 ≤ 0 22
thereby resulting in
𝑑𝐸
𝑑𝑡≤ 0 23
with the equality being valid for
𝑑𝑢𝑖
𝑑𝑡= 0 24
Equation (23) shows that the energy function can never
increase with time which is one of the conditions for a valid
energy function. The second criterion viz. the energy function
must have a lower bound is also satisfied for the circuit of Fig.
1 wherein it may be seen that V1, V2, …, Vn are all bounded (as
they are the outputs of opamps, as given in (21) amounting to
E, as given in (8), having a finite lower bound.
5. SIMULATION RESULTS This sectiondeals with the application of the proposed
network to task of minimizing the objective function
3𝑉12 + 4𝑉1𝑉2 + 5𝑉2
2 25
subject to
𝑉1 − 𝑉2 ≤ −1(26)
𝑉1 + 𝑉2 ≤ 1
The values of resistances acting as the weights on the neurons
are obtained from (17,18). For the purpose of simulation, the
value of K was chosen to be 1 KΩ. Using K = 1 KΩ in
(17,18) gives
Rc11 = Rc12 = Rc21 = Rc22 = K = 1 K, R11 = 1.66 K, R21 = 1 K,
R12 = 1.66 K, R22 = 1 K
For the purpose of PSPICE simulations, the unipolar voltage
comparator was realized using a diode clamp with an opamp
based comparator. The transfer characteristics obtained during
the PSPICE simulations for opamp based bipolar and unipolar
comparators are presented in Fig.6. For the purpose of this
simulation, the LMC7101A CMOS opamp model from the
Orcad library in PSPICE was utilised. The value of βfor this
opamp was measured to be 1.1×104 using PSPICE simulation.
Fig 6: Transfer characteristics for opamp based unipolar
and bipolar comparators
Routine mathematical analysis of (25) yields: V1= – 0.584,
V2= 0.416. The resultant plots of the neuron output voltages
as obtained after PSPICE simulation are presented in Fig.7
from where it can be seen that V(1) = –0.58 V and V(2) =
0.41 V which are very near to the algebraic solution thereby
confirming the validity of the approach. The initial node
voltages were kept as V(1) = –1 mV and V(2) = –10 mV.
Fig 7: Simulation results for the proposed circuit applied
to minimize (25) subject to (26)
6. ISSUES IN VLSI IMPLEMENTATION This section deals with the monolithic implementation issues
of the proposed circuit. The PSPICE simulations assumed that
all operational amplifiers (and diodes) are identical, and
therefore, it is required to determine how deviations from this
assumption affect the performance of the network. Effects of
variations in component values from one neuron to another
were also investigated using Monte-Carlo analysis in PSPICE.
A 10% tolerance with Gaussian deviation profile was put on
the resistances used in the circuit to solve (25). The analysis
was carried out for 100 runs and the Mean Deviation was
found out to be -144.73×10-6 and Mean Sigma (Standard
Deviation) was 0.0119. Offset analysis was also carried out by
incorporating random offset voltages (in the range of 1 mV to
10 mV) to the opamps. The Mean Deviation in this case was
measured to be -143.51×10-6and the Mean Sigma (Standard
Deviation) was 0.012. As can be seen, the effects of
mismatches and offsets on the overall precision of the final
results are in an acceptable range.
In fact, the realization of unipolar comparators by the use of
opamps in the proposed circuit tends to increase the circuit
complexity. The transistor count can be further reduced by
utilising voltage-mode unipolar comparators instead of the
opamp-diode combination. This also suggests that a real, large
scale implementation for solving quadratic programming
problems with high variable counts might be quite different.
International Journal of Computer Applications (0975 – 8887)
Volume 39– No.2, February 2012
48
Alternative realizations based on the differential equations (6)
governing the system of neurons are being investigated. Other
approaches to obtain the tanh(.) non-linearity include the use
of a MOSFET operated in the sub-threshold region [24] and
the use of Current Differencing Transconductance Amplifier
(CDTA) to provide the same nonlinearity in the current-mode
regime [25].
7. CONCLUSION In this paper, a CMOS compatible approach to solve a
quadratic programming problem in n variables subject to
mlinear constraints, which uses nneurons and msynapses is
presented. Each neuron requires one opamp and each synapse
is implemented using a unipolar voltage-mode comparator.
This results in significant reduction in hardware over the
existing schemes. The proposed network was tested on a
sample problem of minimizing a quadratic function in 2
variables and the simulation results confirm the validity of the
approach.
8. REFERENCES [1] Atkociunas, J. 1996. Quadratic programming for
degenerate shakedown problems of bar structures. Mechanics Research Communications, 23(2), 195–206.
[2] Bartlett, R.A., Wachter, A., and Biegler, L.T. 2000. Active set vs. interior point strategies for model predictive control. In Proceedings of the American Control Conference, Chicago, USA, June 2000, 4229–4233.
[3] Maier, G. and Munro, J. 1982. Mathematical programming applications to engineering plastic analysis. Applied Mechanics Reviews, 35, 1631–1643.
[4] Nordebo, S., Zang, Z., and Claesson, I. 2001. A semi-infinite quadratic programming algorithm with applications to array pattern synthesis. IEEE Transactions on Circuits and Systems II: Analog and Digital Signal Processing, 48(3), 225–232.
[5] Schonherr, S. 2002. Quadratic programming in geometric optimization: theory, implementation, and applications. Technical report, Swiss Federal Institute of Technology, Zurich.
[6] Borguet, S. and O. Lonard, O. 2009. A quadratic programming framework for constrained and robust jet engine health monitoring. Progress in Propulsion Physics, 1, 669–692.
[7] Dembo, R.S. and Tulowitzki, U. 1988. Computing equilibria on large multicommodity networks: An application of truncated quadratic programming algorithms. Networks, 18(4), 273–284.
[8] Krogh, A. 2008. What are artificial neural networks? Nature Biotechnology, 26(2), 195–197.
[9] Tank, D.W. and Hopfield, J. 1986. Simple ‘neural’ optimization networks: An A/D converter, signal decision circuit, and a linear programming circuit. IEEE Transactions on Circuits and Systems, 33(5), 533–541.
[10] Kennedy, M.P. and Chua, L.O. 1988. Neural networks for nonlinear programming. IEEE Transactions on Circuits and Systems, 35(5), 554–562.
[11] Maa, C.-Y. andShanblatt, M.A. 1992. Linear and quadratic programming neural network analysis. IEEE Transactions on Neural Networks, 3(4), 580–594.
[12] Chen, Y.-H. and Fang, S.-C. 1998. Solving convex programming problems with equality constraints by neural networks. Computers & Mathematics with Applications, 36(7), 41–68.
[13] Wu, X.-Y., Xia, Y.-S., Li, J., and Chen, W.-K. 1996. A high-performance neural network for solving linear and quadratic programming problems. IEEE Transactions on Neural Networks, 7(3), 643–651.
[14] Xia Y. 1996. A new neural network for solving linear and quadratic programming problems. IEEE Transactions on Neural Networks, 7(6), 1544–1548.
[15] Malek, A. and Alipour, M. 2007. Numerical solution for linear and quadratic programming problems using a recurrent neural network. Applied Mathematics and Computation, 192(1), 27–39.
[16] Rahman, S.A. and Ansari, M.S. 2011. A neural circuitwith transcendental energy function for solving system of linear equations. Analog Integrated Circuits and Signal Processing, 66, 433–440.
[17] Ansari, M.S. and Rahman, S.A. 2010. A DVCC-based non-linear analog circuit for solving linear programming problems. In Proceedings of International Conference on Power, Control and Embedded Systems (ICPCES), Dec 2010, 1–4.
[18] Wang, J. 1992. Recurrent neural network for solving quadratic programming problems with equality constraints. Electronics Letters, 28(14), 1345–1347.
[19] Forti, M. and Tesi, A. New conditions for global stability of neural networks with application to linear and quadratic programming problems. IEEE Transactions on Circuits and Systems I: Fundamental Theory and Applications, 42(7), 354–366.
[20] Tao, Q., Cao, J., and Sun, D. 2001. A simple and high performance neural network for quadratic programming problems. Applied Mathematics and Computation, 124(2), 251–260.
[21] Liu, Q. and Wang, J. 2008. A one-layer recurrent neural network with a discontinuous hard-limiting activation function for quadratic programming. IEEE Transactions on Neural Networks, 19(4), 558–570.
[22] Gould, N.I.M., and Toint, P.L. 2010. A quadratic programming bibliography. Technical report, RAL Numerical Analysis Group, March 2010.
[23] Rahman, S.A., Jayadeva, and S.C. Dutta Roy. 1999. Neural network approach to graph colouring. Electronics Letters, 35(14), 1173–1175.
[24] Newcomb, R.W. and Lohn, J.D. 1998. The handbook of brain theory and neural networks. Chapter: Analog VLSI for neural networks, MIT Press, Cambridge, MA, USA, 86–90.
[25] Ansari, M.S. and Rahman, S.A. 2009. A novel current-mode non-linear feedback neural circuit for solving linear equations. In Proceedings of International Conference on Multimedia, Signal Processing and Communication Technologies, 284–287.
International Journal of Computer Applications (0975 – 8887)
Volume 39– No.16, February 2012
31
A Non–linear Feedback Neural Network for
Graph Coloring
Mohd. Samar Ansari Department of Electronics
Engineering Aligarh Muslim University
Aligarh, India
Syed Atiqur Rahman Department of Electronics
Engineering Aligarh Muslim University
Aligarh, India
Syed Javed Arif Department of Electronics
Engineering Aligarh Muslim University
Aligarh, India
ABSTRACT
A feedback neural network for solving graph coloring
problem is presented. The circuit has an associated
transcendental energy function that ensures fast convergence
to the exact solution. Hardware and PSPICE simulation
results on random and benchmark problems have been
presented. Test results are compared with existing techniques
for graph coloring to show that the proposed neural network
model provides a significant reduction in the number of colors
while enjoying a simple and efficient circuit implementation.
General Terms
Neural Networks, Graph Coloring.
Keywords
Neural network applications, Neural network hardware, Non-
linear circuits, Graph theory,Dynamical Systems, Non-Linear
Synapse,Feedback Networks.
1. INTRODUCTION The graph coloring problem (GCP) has a large number of
applications such as register allocation in digital computers,
frequency channel assignment in mobile communication,
and layer assignment in VLSI design [1-3]. In GCP,
values or colors are assigned to different nodes of a
graph such that no two adjacent nodes get the same
color, and the number of colors are also minimized. The
Hopfield Neural Network (HNN) and its
modifications have been applied to solve GCP, but the
implementation requires N2 neurons and at most N4
weights for coloring a graph of N nodes [4]. This puts
severe constraints on the feasibility of HNN realization in
hardware. Another feedback neural network architecture
which uses non-linear synapses, and has a transcendental
Energy function, requires N neurons and at most N2
interconnections for solving a problem with N nodes [4]. In
this paper, a neural circuit is presented which modifies the
Energy function of [4] to eliminate those local minima that
correspond to solutions with large number of colors. It is seen
that the number of colors are reduced thereby providing a
significant improvement in solution quality. Moreover, the
circuit is suitable for VLSI implementation as compared to an
existing scheme which employs one coupled relaxation
oscillator for each node in the graph to be colored [5].
2. PROPOSED CIRCUIT FOR GRAPH
COLORING Fig. 1 shows thei-th neuron of proposed network. In the
proposed circuit, output voltages of different neurons
represent the colors of different nodes. Ci and ri denote the
internal capacitance and resistance of the i-th neuron
respectively, ui is the internal state and Rii is the self–feedback
resistance of i-th neuron. The output of other neurons Vj (j =
1, 2…N) are connected to the input of i-th neuron through
unipolar comparators.
ri Ci
Rii
x i1
xij
xiN
Vi
V1
Vj
VN
Ri1
Ri j
RiN
ui
Vi
Ni
Fig 1: i-th neuron of the proposed neural network for
graph coloring
The self – feedback resistance (Rii) is given by
𝑅𝑖𝑖 = 𝑅𝑐
∆ (1)
whereRc is a constant resistance and ∆ is the maximum degree
of the graph. Next, we define a constant aij such that
𝑎𝑖𝑗 = 1; 𝑖 − 𝑡ℎ 𝑛𝑜𝑑𝑒 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑒𝑑 𝑡𝑜 𝑗 − 𝑡ℎ 𝑛𝑜𝑑𝑒0; 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
(2)
Using (2), the feedback resistance of the i-th neuron is given
by
𝑅𝑖𝑗 = 𝑅𝑐
𝑎𝑖𝑗 (3)
From fig. 1, xij can be written as
𝑥𝑖𝑗 =𝑉𝑚
2 𝑡𝑎𝑛ℎ 𝛽 𝑉𝑗 − 𝑉𝑖 − 1 (4)
where β is the open-loop gain of the comparator (practically
very high) and ±Vm are the voltage levels of the comparator
output.
Node equation for node „Ni‟ gives the equation of motion of
the i-th neuron in the state space as
𝐶𝑖𝑑𝑢 𝑖
𝑑𝑡=
𝑥𝑖𝑗
𝑅𝑖𝑗
𝑁𝑗=1𝑗≠𝑖
+ 𝑉𝑖
𝑅𝑖𝑖 −
𝑢 𝑖
𝑅𝑖(5)
International Journal of Computer Applications (0975 – 8887)
Volume 39– No.16, February 2012
32
where
1
𝑅𝑖=
1
𝑅𝑖𝑗
𝑁𝑗=1𝑗≠𝑖
+ 1
𝑅𝑖𝑖+
1
𝑟𝑖(6)
Moreover, it can be shown that the Energy Function „E‟ of the
network in Fig. 1 is
𝐸 = 1
2
𝑉𝑖2
2𝑅𝑖𝑖
𝑁𝑖=1 −
𝑉𝑚
4𝛽𝑅𝑐
𝑎𝑖𝑗𝑁𝑗=1𝑗≠𝑖
𝑁𝑖=1 ln cosh 𝛽 𝑉𝑗 − 𝑉𝑖 −
𝑉𝑚
2𝑅𝑐
𝑎𝑖𝑗𝑁𝑗=1𝑗≠𝑖
𝑁𝑖=1 𝑉𝑖 −
1
𝑅𝑖 𝑢𝑖𝑉𝑖
0𝑑𝑉𝑁
𝑖=1 (7)
Fig 2: Plot of the second and third terms in equation (7)
The last term in (7) is usually neglected for high values of the
open-loop gain of the opamp used to realize the neurons. The
first term on the right hand side of (7) is quadratic which tries
to minimize the number of colors. The second term has got a
negative sign. Therefore, the energy function (E) will be
minimized if second term is maximized. This happens when
the voltages corresponding to connected nodes in a graph are
far away from each other. The first two terms on the right
hand side are balancing each other to color a graph properly.
The third term also contributes to lowering of number of
different colors by eliminating all those local minima in the
energy function for which node voltages are negative. The
combined effect of the second and third terms in (7) for a
typical 2-node problem is shown in Fig. 2.
3. SIMULATION RESULTS The proposed network was tested for various random graphs
as well as a standard benchmark problem for graph coloring,
myciel3.col [6]. Breadboard implementation as well as
PSPICE simulation was performed. The results of the tests are
given in table 1 from which it is seen that the proposed
network gives a solution to all the problems tested and in all
the cases the solution is very near to the chromatic number of
the graph. Simulation runs of the proposed network for the
myciel3.col benchmarking problem returned the best solution
of 5 colors which is very near to the chromatic number for
myciel3.col, that being 4. The performance of the proposed
network was also compared with a network proposed earlier
[4]. Table 2 gives this performance comparison. It is evident
that substantial improvements have been achieved both in the
best and the average solutions for most examples.
Table1.Hardware and PSPICE simulation test results for the proposed network
International Journal of Computer Applications (0975 – 8887)
Volume 39– No.16, February 2012
33
4. DISCUSSION This section deals with the monolithic implementation issues
of the proposed circuit. The PSPICE simulations assumed that
all operational amplifiers are identical, and therefore, analysis
is required to determine how deviations from this assumption
affect the performance of the network. While the component
values used for the capacitors and resistors affect the
convergence time, in a real implementation, they would also
need to be chosen keeping in mind limitations of the active
devices to source or sink currents. Effects of variations in
component values from one neuron to another also need to be
investigated. In fact, the drawbacks of the use of opamps in
the proposed circuit suggest that a real, large scale
implementation of the circuit for colouring a graph with large
number of nodes might be quite different. Alternative
realizations based on the differential equations (5) governing
the system of neurons are being investigated. An
implementation based on operational transconductance
amplifiers (OTAs) would appear to be one viable alternative;
similarly, circuits using MOS transistors operating in the
subhreshold regime would be another. In the latter case, the
tanh(.) nonlinearity is easily obtained because of the
transistor‟s current–voltage characteristic [7] and is expected
to be the focus of future work. However, the present
manuscript focuses mainly on the principle of such a network,
and a VLSI implementation is beyond the scope of this paper.
5. CONCLUSION A feedback neural network based architecture to solve graph
coloring problem is presented. Test results of hardware
implementation and PSPICE simulation for five different
graphs including the myciel3.col benchmark problem are
presented. Further, PSPICE simulations of the proposed
circuit show improvements in both the minimum and the
average number of colors over its predecessor.
6. REFERENCES [1] Zeitlhofer, T. and Wess, B. 2004. A Comparison of
Graph Coloring Heuristics for Register Allocation based
on Coalescing in Interval Graphs. Proc. ISCAS 04, vol.
IV, 529 – 532.
[2] El-Fishawy, N.A., Hadhood, M. M., Elnoubi, S., and El-
Sersy, W.2000. A modified Hopfield neural network
algorithm for cellular radio channel assignment. Proc.
TENCON 2000, 2, 213 – 216.
[3] Blas, A.D., Jagota, A., and Hughey, R. 2002. Energy
Function – Based Approaches to Graph Coloring. IEEE
Trans. On Neural Networks, 13, 81–91.
[4] Rahman, S.A., Jayadeva and Dutta Roy, S.C. 1999.
Neural network approach to graph colouring. Electronics
Letters, 35(14), 1173–1175.
[5] Wu, C.W. 1998. Graph Coloring via Synchronization of
Coupled Oscillators. IEEE Trans. Circuits Syst. I, 45(9),
974–978.
[6] Lui, J., Zhong, W., and Jiao, L.2006. Comments on “The
1993 DIMACS Graph Coloring Challenge” and “Energy
Function-Based Approaches to Graph Coloring” IEEE
Trans. On Neural Networks, 17(2), 533.
[7] Newcomb, R.W. and Lohn, J.D.1998. Analog VLSI for neural networks in The Handbook of Brain Theory and Neural Networks. Cambridge, MA: The MIT Press.
Table2.Performance comparison of the proposed and existing networks
A Non – Linear Feedback Neural Network for Solving Quadratic
Programming Problems
MOHD. SAMAR ANSARI
1,* AND S.A. RAHMAN2
1,2
Department of Electronics Engineering, Aligarh Muslim University, Aligarh, India
Emails: [email protected], [email protected]
2
Abstract- This paper presents a neural circuit for solving quadratic programming problems. The objective is to minimize
a quadratic cost function subject to linear constraints. The proposed circuit employs non-linear feedback, in the form of
unipolar comparators, to introduce transcendental terms in the energy function ensuring fast convergence to the
solution. PSPICE simulation results are presented for a chosen optimization problem and are found to agree with the
algebraic solution.
Keywords: Neural Networks, Non – Linear circuits, Feedback Neural Networks, Quadratic Programming, Dynamical
Systems
1. Introduction
In the discipline of constrained optimization, problems with nonlinear objective functions are usually
approximated by a second-order (quadratic) system and solved approximately by a standard quadratic
programming technique. Traditional methods for solving quadratic programming problems typically involve
an iterative process, but long computational time limits their usage. There is an alternative approach to
solution of this problem. It is to exploit the artificial neural networks (ANN's) which can be considered as an
analog computer relying on a highly simplified model of neurons [1]. ANN's have been applied to several
classes of constrained optimization problems and have shown promise for solving such problems more
effectively. For example, the Hopfield neural network has proven to be a powerful tool for solving some of
the optimization problems. Tank and Hopfield first proposed a neural network for solving
mathematical programming problems, where a linear programming problem was mapped into a closed-
loop network [2]. Later, the dynamical approach was extended for solving quadratic programming
problems. Over the past two decades several neural-network architectures for solving quadratic
programming problems have been proposed by Kennedy and Chua [3], Maa and Shanblatt [4], Chen and
Feng [5], Wu et al. [6], Xia [7]. These networks depend on the network parameters or use expensive analog
multipliers. More recently, Malek and Alipour [8] proposed a recurrent neural network that is able to solve
quadratic programming problems. This architecture is advantageous in the sense that there is no need to set
network parameters. Moreover, the number of analog multipliers is reduced. This reduces the hardware
complexity to a certain extent. However, the network still has a complex hardware in the sense that the
main processing element (alpha) would be cumbersome to implement in hardware. In this paper, we present
a hardware solution to the problem of solving a quadratic programming problem subject to linear constraints.
The proposed architecture uses non-linear feedback which leads to a new energy function that involves
transcendental terms. This transcendental energy function is fundamentally different from the standard
quadratic form associated with Hopfield network and its variants. The hardware complexity of the proposed
circuit compares favourably with the existing hardware implementations.
The remainder of this paper is arranged as follows: Section – 2 outlines the basic problem and the
mathematical formulation on which the development of the proposed network will be based. Section – 3
contains the circuit implementation of the proposed network for a set of sample problem in two variables.
SPICE simulation results of the proposed circuit are also presented. Issues that are expected to arise in actual
monolithic implementations are discussed in Section – 4. Concluding remarks are presented in Section – 5.
2. Proposed Circuit
Let the function to be minimized be
𝐹 = 𝐶𝑖𝑗𝑉𝑖𝑉𝑗𝑛𝑗=1
𝑛𝑖=1 (1)
Subject to
a11 a12 … a1n
a21 a22 … a2n
⋮ ⋮ … ⋮am1 am2 … amn
𝑉1
𝑉2
⋮𝑉𝑛
≤
b1
b2
⋮bm
(2)
Where V1, V2,..., Vn are the variables and aij, cij and bi are constants. The proposed neural-network based
circuit to minimize the quadratic function given in (1) in accordance with the constraints of (2) is presented in
fig 1. As can be seen from fig. 1, individual equations from the set of equations to be solved are passed
through non-linear synapses which are realized using comparators. The comparator outputs can be modelled
by equations (3) given below.
𝑥1 =𝑉𝑚
2 𝑡𝑎𝑛ℎ 𝛽 𝑎11𝑉1 + 𝑎12𝑉2 + ⋯ + 𝑎1𝑛𝑉𝑛 − 𝑏1 + 1
𝑥2 =𝑉𝑚
2 𝑡𝑎𝑛ℎ 𝛽 𝑎21𝑉1 + 𝑎22𝑉2 +⋯ + 𝑎2𝑛𝑉𝑛 − 𝑏2 + 1 (3)
⋮
𝑥𝑚 =𝑉𝑚
2 tanh 𝛽 𝑎𝑚1𝑉1 + 𝑎𝑚2𝑉2 +⋯ + 𝑎𝑚𝑛 𝑉𝑛 − 𝑏𝑚 + 1
where 𝛽 is the open-loop gain of the comparator (practically very high), ±Vm are the output voltage levels of
the comparator and V1, V2,..., Vn are the neuron outputs. The outputs of the comparators are fed to neurons
having weighted inputs. The neurons are realized by using opamps and the weights are implemented using
transconductance amplifiers. The currents arriving to the neuron from various synapses get added up at the
input of the neuron. Rpi and Cpi are the input resistance and capacitance of the opamp corresponding to the
i-th neuron. These parasitic components are included to model the dynamic nature of the opamp.
Vi
Vout
+
_
Vi
Vout
+
_
+
_
Vi
b1
b2
x1
x2
Rpi Ci
Vi
Vout
+
_bnxm
G1
G2
Gn
a11V1+ a12V2 + ...+a1nVn
a21V1+ a22V2 +...+ a2nVn
am1V1+ am2V2 + ...+amnVn
g1i x1
g2i x2
gmi xm
To other
neurons
V1 V2 Vn
R1i R2i Rni
Ni
Fig. 1 i-th neuron of the proposed feedback neural network circuit to solve a quadratic programming problem in n-
variables with m- linear constraints
Applying node equations for node ‘Ni’ we can write the equations of motion of the various neurons. For the
first neuron,
𝐶1
𝑑𝑢1
𝑑𝑡= 𝑥1𝑔𝑐11 + 𝑥2𝑔𝑐21 + ⋯ + 𝑥𝑚𝑔𝑐𝑚1 +
𝑉1
𝑅11+
𝑉2
𝑅21+ ⋯ +
𝑉𝑛𝑅𝑛1
− 𝑢1 𝑔𝑐11 + 𝑔𝑐21 + ⋯ + 𝑔𝑐𝑚1 +1
𝑅11+
1
𝑅21+ ⋯+
1
𝑅𝑛1+
1
𝑅𝑝1 (4)
Simplifying, we get
𝐶1
𝑑𝑢1
𝑑𝑡= 𝑥1𝑔𝑐11 + 𝑥2𝑔𝑐21 + ⋯ + 𝑥𝑛𝑔𝑐𝑚1 +
𝑉1
𝑅11+
𝑉2
𝑅21+ ⋯ +
𝑉𝑛𝑅𝑛1
− 𝑢1 1
𝑅1 (5)
Where
1
𝑅1=
1
𝑅11+
1
𝑅21+ ⋯ +
1
𝑅𝑛1+ 𝑔𝑐11 + 𝑔𝑐21 + ⋯ + 𝑔𝑐𝑚1 +
1
𝑅𝑝1 (6)
Similarly, we have for the second neuron
𝐶2
𝑑𝑢2
𝑑𝑡= 𝑥1𝑔𝑐12 + 𝑥2𝑔𝑐22 + ⋯+ 𝑥𝑛𝑔𝑐𝑚2 +
𝑉1
𝑅12+
𝑉2
𝑅22+ ⋯ +
𝑉𝑛𝑅𝑛2
− 𝑢2 1
𝑅2 (7)
Where
1
𝑅2=
1
𝑅12+
1
𝑅22+ ⋯ +
1
𝑅𝑛2+ 𝑔𝑐12 + 𝑔𝑐22 + ⋯ + 𝑔𝑐𝑚2 +
1
𝑅𝑝2 (8)
And, for the n-th neuron we get,
𝐶𝑛
𝑑𝑢𝑛
𝑑𝑡= 𝑥1𝑔𝑐1𝑛 + 𝑥2𝑔𝑐2𝑛 + ⋯+ 𝑥𝑛𝑔𝑐𝑚𝑛 +
𝑉1
𝑅1𝑛+
𝑉2
𝑅2𝑛+ ⋯ +
𝑉𝑛𝑅𝑛𝑛
− 𝑢𝑛 1
𝑅𝑛 (9)
Where
1
𝑅𝑛=
1
𝑅1𝑛+
1
𝑅2𝑛+ ⋯ +
1
𝑅𝑛𝑛+ 𝑔𝑐1𝑛 + 𝑔𝑐2𝑛 + ⋯ + 𝑔𝑐𝑚𝑛 +
1
𝑅𝑝𝑛 (10)
Where ui is the internal state of the i-th neuron and gcji is the voltage-to-current conversion factor of the j-th
transconductance block for the i-th output current. As is shown later in this section, these weights are
governed by the entries in the coefficient matrix of (2).
Moreover, the network in fig. 1 can be associated with an Energy Function ‘E’ of the form
𝐸 = 𝐶𝑖𝑗𝑉𝑖𝑉𝑗
𝑛
𝑗=1
𝑛
𝑖=1
+ 𝑉𝑚2
𝑎𝑖𝑗𝑉𝑗
𝑛
𝑗=1
𝑚
𝑖=1
+ 𝑉𝑚2𝛽
ln coshβ (𝑎𝑖𝑗𝑉𝑗
𝑛
𝑗=1
− 𝑏𝑖)
𝑚
𝑖=1
(11)
This expression of the Energy Function can be written in a slightly different (but more illuminating) form as
𝐸 = 𝐶𝑗𝑖𝑉𝑖𝑉𝑗
𝑛
𝑗=1
𝑛
𝑖=1
+ 𝑃1 + 𝑃2 + ⋯ + 𝑃𝑚 (12)
Where P1, P2, ..., Pm are the penalty terms. The penalty terms can separately be given as
𝑃1 = 𝑉𝑚2
𝑎11𝑉1 + 𝑎12𝑉2 + ⋯ + 𝑎1𝑛𝑉𝑛 +𝑉𝑚2𝛽
ln coshβ( 𝑎11𝑉1 + 𝑎12𝑉2 + ⋯ + 𝑎1𝑛𝑉𝑛 − 𝑏1) (13)
Similarly, the remaining penalty functions can be written, as given below.
𝑃2 = 𝑉𝑚2
𝑎21𝑉1 + 𝑎22𝑉2 + ⋯ + 𝑎2𝑛𝑉𝑛 +𝑉𝑚2𝛽
ln coshβ(𝑎21𝑉1 + 𝑎22𝑉2 + ⋯ + 𝑎2𝑛𝑉𝑛 − 𝑏2) (14)
And finally,
𝑃𝑚 = 𝑉𝑚2
𝑎𝑚1𝑉1 + 𝑎𝑚2𝑉2 + ⋯ + 𝑎𝑚𝑛 𝑉𝑛 +𝑉𝑚2𝛽
ln coshβ( 𝑎𝑚1𝑉1 + 𝑎𝑚2𝑉2 + ⋯+ 𝑎𝑚𝑛 𝑉𝑛 − 𝑏𝑚 ) (15)
Obtaining a partial differentiation of the combined penalty term, P w. r. t. to V1 we get
𝜕𝑃
𝜕𝑉1=
𝜕𝑃1
𝜕𝑉1+
𝜕𝑃2
𝜕𝑉1+ … +
𝜕𝑃𝑚𝜕𝑉1
(16)
𝜕𝑃
𝜕𝑉1=
𝑉𝑚2
𝑎11 + 𝑎21 + ⋯ + 𝑎𝑚1
+ 𝑎11 tanh β( 𝑎11𝑉1 + 𝑎12𝑉2 + ⋯ + 𝑎1𝑛𝑉𝑛 − 𝑏1)
+ 𝑎21 tanh β( 𝑎21𝑉1 + 𝑎22𝑉2 + ⋯ + 𝑎2𝑛𝑉𝑛 − 𝑏2) + ⋯ +𝑎𝑚1 tanh β( 𝑎𝑚1𝑉1 + 𝑎𝑚2𝑉2 + ⋯
+ 𝑎𝑚𝑛 𝑉𝑛 − 𝑏𝑚 ) (17)
Which can be simplified to,
𝜕𝑃
𝜕𝑉1= 𝑎11𝑥1 + 𝑎21𝑥2 + ⋯ + 𝑎𝑚1𝑥𝑚 (18)
Using the above relations to find the derivative of the Energy Function w. r. t. to V1 we have
𝜕𝐸
𝜕𝑉1=
𝜕
𝜕𝑉1 𝐶𝑗𝑖𝑉𝑖𝑉𝑗
𝑛
𝑗=1
𝑛
𝑖=1
+ 𝜕𝑃
𝜕𝑉1 (19)
Which yields
𝜕𝐸
𝜕𝑉1= 𝑐11𝑉1 + 𝑐12𝑉2 + ⋯ + 𝑐1𝑛𝑉𝑛 + 𝑎11𝑥1 + 𝑎21𝑥2 + ⋯ + 𝑎𝑚1𝑥𝑚 (20)
Also, if ‘E’ is the Energy Function, it must satisfy the following condition *9].
𝜕𝐸
𝜕𝑉𝑖= 𝐾𝐶𝑖
𝑑𝑢𝑖
𝑑𝑡 (21)
Where ‘K’ is a constant of proportionality and has the dimensions of resistance. Equation (x) applied to the
first neuron results in
𝑔𝑐11 =𝑎11
𝐾
𝑔𝑐21 =𝑎21
𝐾 (22)
⋮
𝑔𝑐𝑚1 =𝑎𝑚1
𝐾
A similar comparison of the remaining partial fractions for the remaining variables yields the following:
𝑔𝑐11 𝑔𝑐12
𝑔𝑐21 𝑔𝑐22
… 𝑔𝑐1𝑛
… 𝑔𝑐2𝑛
⋮ ⋮𝑔𝑐𝑚1 𝑔𝑐𝑚2
… ⋮… 𝑔𝑐𝑚𝑛
=1
𝐾
𝑎11 𝑎12
𝑎21 𝑎22
… 𝑎1𝑛
… 𝑎2𝑛
⋮ ⋮𝑎𝑚1 𝑎𝑚2
… ⋮… 𝑎𝑚𝑛
(23)
And,
𝑅11 𝑅12
𝑅21 𝑅22
… 𝑅1𝑛
… 𝑅2𝑛
⋮ ⋮𝑅𝑛1 𝑅𝑛2
… ⋮… 𝑅𝑛𝑛
= 𝐾
1
𝑐11 1
𝑐12
1𝑐21
1𝑐22
… 1𝑐1𝑛
… 1𝑐2𝑛
⋮ ⋮1
𝑐𝑛1 1
𝑐𝑛2
… ⋮
… 1𝑐𝑛𝑛
(24)
3. Simulation Results
This section deals with the application of the proposed network to task of minimizing the objective function
3𝑉12 + 4𝑉1𝑉2 + 5𝑉2
2 (25) subject to
𝑉1 − 𝑉2 ≤ −1 (26)
𝑉1 + 𝑉2 ≤ 1 The values of resistances acting as the weights on the neurons are obtained from (24, 25). For the purpose of
simulation, the value of ‘K’ was chosen to be 1 KΩ. Using K = 1 KΩ in (23, 24) gives
gc11 = gc12 = gc21 = gc22 = 1/K = 1
R11 = 1.66 KΩ, R21 = 1 KΩ, R12 = 1.66 KΩ, R22 = 1 KΩ
For the purpose of HSPICE simulations, the voltage comparator and the multi-output transconductance
amplifier for each neuron were realized as a single block using a slightly modified circuit of the differential-
input, high-gain, active-loaded transconductance amplifier [10] shown in fig. 2. Standard BSIM3 0.35µm
parameters were used for the purpose of simulations. The supply voltages were set to VDD = 15V, VSS = 0V
and VBB = 1V. The aspect ratios of the NMOS and PMOS transistors were taken to be 1.4µm/1.4µm and
2.8µm/1.4µm respectively. Further, to get the output currents according to (23) the W/L ratios of transistors
M6, M7, M8 and M9 were set to provide the required current scaling. Choosing K=1 and using (23) results in
Iout1 = Iout2 for G1 block and Iout1 = Iout2 for the G2 block.
M2
M1
M3
M4
M5
M6
M7
VDD
VSS
VBB
V2
V1 I
out1
M8
M9
Iout2
Fig. 2 CMOS implementation of the comparator and transconductance blocks
Fig. 3 Simulation results for the proposed circuit applied to minimize (25) subject to (26)
Routine mathematical analysis of (25) yields: 𝑉1= -0.584, 𝑉2= 0.416. The results of PSPICE simulation are
presented in fig. 3. From the plots of the neuron output voltages, it can be seen that V(1) = -0.58 V and V(2) =
0.41 V which are very near to the algebraic solution thereby confirming the validity of the approach.
4. Issues in actual implementation
This section deals with the monolithic implementation issues of the proposed circuit. The PSPICE simulations
assumed that all operational amplifiers (and transconductance amplifiers) are identical, and therefore,
analysis is required to determine how deviations from this assumption affect the performance of the
network. Effects of variations in component values from one neuron to another also need to be investigated.
Moreover, for obtaining weighted (scaled) current outputs use was made of the (W/L) ratios of the
transistors M6, M7, M8 and M9 of fig. 2. This results in a particular circuit implementation ‘hard-wired’ to
solve a given quadratic programming problem. Methods to scale the currents by a mechanism external to
the circuit need to be explored to impart flexibility to the circuit. Alternative realizations based on the
differential equations (9) governing the system of neurons are also being considered. Other approaches to
obtain the tanh(.) nonlinearity include the use of a MOSFET operated in the subthreshold region [10] and the
use of Current Differencing Transconductance Amplifier (CDTA) to provide the same nonlinearity in the
current-mode regime [11].
5. Conclusion
In this paper we have described a CMOS compatible approach to solve a quadratic programming problem in
n variables subject to m linear constraints, which uses n-neurons and m-synapses. Each neuron requires one
opamp and each synapse is implemented using a differential transconductance element. This results in
significant reduction in hardware over the existing schemes [3-8]. The proposed network was tested on a
sample problem of minimizing a quadratic function in 2 variables and the simulation results confirm the
validity of the approach.
REFERENCES
[1] J.J. Hopfield, D.W. Tank, (1985) ‘Neural’ computation of decisions optimization problems, Biological
Cybern., 52, 141-152,.
[2] D.W. Tank, J.J. Hopfield, (1986) Simple neural optimization networks: an A/D converter, signal decision
network, and linear programming circuit, IEEE Trans. Circ. Syst. CAS, 33, 533–541.
[3] M.P. Kennedy, L.O. Chua, (1988) Neural networks for nonlinear programming, IEEE Trans. Circ. Syst.,
35, 554–562.
[4] C.Y. Maa, M.A. Shanblatt, (1992) A two-phase optimization neural network, IEEE Trans. Neural
Network 3 (6), 580 – 594.
[5] Y.-H. Chen, S.-C. Fang, (1998) Solving convex programming problems with equality constraints by
neural networks, Comput. Math. Appl., 36 (7), 41–68.
[6] X.Y. Wu, Y.S. Xia, J. Li, W.K. Chen, (1996) A high performance neural network for solving linear and
quadratic programming problems, IEEE Trans. Neural Networks, 7 (3), 643–651.
[7] Y. Xia, (1996) A new neural network for solving linear and quadratic programming problems, IEEE
Trans. Neural Networks 7 (6), 1544–1547.
[8] A. Malek, M. Alipour, (2007), Numerical solution for linear and quadratic programming problems using a
recurrent nueral network, Appl. Math. Computation, 192, 27 – 39.
[9] S. A. Rahman, Jayadeva, S. C. Dutta Roy, (1999) Neural network approach to graph colouring,
Electronics Letters, 35, 1173 – 1175.
[10] R. W. Newcomb, J. D. Lohn, (1998) Analog VLSI for neural networks, The Handbook of Brain Theory
and Neural Networks. Cambridge, MA: The MIT Press.
[11] M. S. Ansari, S. A. Rahman. (2009) A Novel Current-mode Non-Linear Feedback Neural Circuit for
Solving Linear Equations, Proc. IMPACT – 2009, Aligarh, India, 284 – 287.
In: CMOS Technology ISBN: 978-1-61761-325-8
Editor: Min-jun Kwon © 2010 Nova Science Publishers, Inc.
Chapter 3
LINEAR AND NON-LINEAR APPLICATIONS
OF CMOS DVCC
Sudhanshu Maheshwari, Mohd. Samar Ansari
and Syed Atiqur Rahman Department of Electronics Engineering,
Aligarh Muslim University, Aligarh, India
ABSTRACT
The differential voltage current conveyor (DVCC) has recently emerged as a
versatile building block for the realization of a variety of analog signal processing
functions including, but not limited to, analog filters, amplifiers, integrators, etc. Most of
these applications can be classified as linear although non-linear applications like the
voltage-mode comparator, function generators and oscillators are also reported in the
technical literature.
This chapter presents several new and traditional applications of DVCC including
both linear and non-linear ones. Amongst the linear applications, voltage-mode and
current-mode analog filters, amplifiers and integrators are reviewed. Similarly, amongst
the available non-linear ones, oscillators and negative resistance converter are reviewed
whereas DVCC-based digital logic gates, linear equation solver and circuits for solving
linear & quadratic programming problems are proposed. Real device simulations on some
of the newly presented circuits are also given. The material presented is intended to
explore full potential of DVCC covering a broad range of practical circuit applications.
I. INTRODUCTION
Over the past few decades, a variety of active building blocks (including, but not limited
to, Current Conveyors, Current-Controlled Conveyors, Operational Transconductance
Amplifiers, Current Differencing Transconductance Amplifier, Dual-X Current Conveyor,
etc.) and their applications have been added to the technical literature [1-5]. Amongst these,
Sudhanshu Maheshwari, Mohd. Samar Ansari and Syed Atiqur Rahman
2
the Second Generation Current Conveyor (CCII) has received the most attention of circuit
designers. This is because of the fact that the CCII proves to be a versatile building block that
can be used to implement a variety of high-performance circuits which are simple to
construct. The CCII has a disadvantage that only one of the input terminals exhibits a high
input impedance (Y terminal). This disadvantage becomes evident when the CCII is required
to handle differential signals, as in the case of an instrumentation amplifier. Differential
signal processing also has the advantage of better rejection of common-mode signals. The
first step towards removing the said disadvantage came in 1989 when Pal [6] proposed a new
form of the CCII and called it the modified CCII. Pal‟s modified CCII and the resultant
DVCC II are presented in Figure 1 below.
Although enjoying differential inputs, the DVCC II of Figure 1 was not taken up very
enthusiastically by the research fraternity. The DVCC remained a neglected active building
block till 1997 when Elwan and Soliman [7] came up with a novel realization. This DVCC
implementation, shown in Figure 2, utilized the dominant monolithic integration technology
viz. CMOS.
Elwan and Soliman [7] were also the first to give a formal matrix equation (1) describing
the functionality of the five port DVCC.
[ ]
[
]
[ ]
(1)
Z
Z
Y1
Y2
X
CC II
±
+
-
Y
Z
Z
Y1
Y2
X
DVCC II
±
Figure 1. Modified Current Conveyor and its symbolic equivalent [6]
Y2 Y1 X Z1 Z2
Vb
Vb1
VSS
VDD
M2M1 M3 M4
M5 M6
M7 M8 M9 M10
M11
M12
M13
M14 M15 M16
M17 M18
Figure 2. CMOS implementation of the DVCC [7]
Linear and Non-Linear Applications of CMOS Dvcc
3
VCM
Y1Y2
VB1
VB2
VB1
VBX
VB2 VB3
X Z
VDD
VSS
M1 M2 M3 M4
M5 M6 M7 M8
M9M10
M11
M13
M12
M14
M15
M17
M16
M18
M19
M20
M21
M22 M24
M23
Figure 3. Low-Voltage Low-Power CMOS DVCC [9]
Several new and improved CMOS implementations were reported in the technical
literature after the circuit of Figure 2. These included a DVCC realization based on Quasi-
Floating Gate (QFG) transistors proposed by Moradzadeh and Azhari [8]. Mahmoud
presented a low-voltage low-power implementation of the DVCC which is reproduced in
Figure 3 for reference [9].
Another popular version of the DVCC, which is a slightly modified form of the circuit of
Figure 2, has been used in [10] and is reproduced in Figure 4. Other noteworthy DVCC
realizations can be attributed to Hu et al. [11], Siripruchyanan and Jaikla [12] (a bipolar
implementation) and, more recently, Hassan and Mahmoud [13].
Y1Y2
VBB
M1 M4M2 M3
M5 M6
M9 M10
M7
M11
M8
M12
XZ+
VSS
VDD
M16
Z-
M17 M18
M15M14M13
Figure 4. Improved CMOS implementation of DVCC [10]
Sudhanshu Maheshwari, Mohd. Samar Ansari and Syed Atiqur Rahman
4
Y1
Y2
Z+
Z-
DVCC
X
Y2
Y1
IIN Iout1
Iout2
Figure 5. DVCC based basic current processing block [15]
II. SURVEY OF SOME EXISTING DVCC BASED CIRCUITS
The literature is replete with both linear and non-linear applications of DVCC [6-27].
Amongst the linear DVCC-based circuits that are reported, some are discussed here.
A. Amplifiers
A voltage-mode amplifier using DVCC was presented in [14]. Another amplifier circuit
that needs a mention is the versatile building block shown in Figure 5 [15]. It can be made to
work as a current-mode summing amplifier by proper selection of the admittances Y1 and Y2.
For instance, if the admittances are Y1 = 1/ R1 and Y2 = 1/ R2, the current-mode amplifier can
be realized with a gain equal to (R1/ R2).
A DVCC-based instrumentation amplifier was presented in [13] and is reproduced in
Figure 6 below. The circuit takes a differential input voltage, multiplies it with a gain, and
produces a single-ended output voltage. The relation between the output voltage and the
differential input voltage was shown to be
(2)
Y2
Y1
DVCC
X
Z+
R1
VOUT
R2
+
-VIN-DIFF
Figure 6. DVCC based Instrumentation Amplifier [13]
Linear and Non-Linear Applications of CMOS Dvcc
5
B. Integrators
A voltage-mode integrator using DVCC was presented in [14]. A current-mode integrator
circuit was obtained from the basic current processing block, shown in Figure 6, in [15] by a
judicious selection of the admittances Y1 and Y2. By choosing Y1 = sC1 and Y2 = 1/ R2 and
the transfer function was shown to be
(3)
C. Filters with Bi-Linear Transfer Functions
Another application area where the DVCC has been put to good use is the design of
voltage-mode and current-mode analog filters. First, a current-mode first-order all-pass filter
section is discussed. The circuit of Figure 7, attributed to Minaei [16], presents the DVCC
based first order AP (all-pass) filter. The advantage of using a DVCC can be understood by
the fact that due to the differential input at Y-terminals, a very simple all-pass filter
realization is obtained. Another noteworthy point is that the circuit is cascadable and employs
only grounded passive components. This makes the circuit attractive for monolithic
integration. Selecting R2=2R1, the current transfer function for the AP filter of Figure 7 [16] is
given in (4) below. By using (4), the gain and phase response of the filter were shown to be
those given in (5) and (6).
A current-mode all-pass section using a new „modified‟ DVCC was reported in [17] and
is reproduced in Figure 8. The DVCC utilized in this case had two Z+ outputs and no Z-
output. Moreover, one of the Z+ stages had a gain of 2. This was achieved by tweaking the
aspect ratios of the transistors comprising the second Z+ stage.
(4)
(5)
( ) ( ) (6)
Y1
Y2
Z+
Z-
DVCC
X
Iout
Iout
R1
R2C1
IIN
IIN
R3
Figure 7. First-Order Dual-Input All-Pass Filter using a Single DVCC [16]
Sudhanshu Maheshwari, Mohd. Samar Ansari and Syed Atiqur Rahman
6
Y2 Z+
2Z+
DVCC
X
RX
Iout
Iin
Iin
C
Figure 8. Current-mode All-Pass Filter using a Modified DVCC [17]
Assuming ideal DVCC operation, the current transfer gain for the circuit of Figure 8 [17]
is reproduced below.
(
⁄ )
( ⁄ ) (7)
Equation (7) is the standard transfer function of an all-pass filter. The circuit of Figure 8
therefore provides a constant unity gain at all frequencies and frequency dependent phase
with a value
( ) ( ) (8)
The circuit [17] with only two passive components, both grounded, is quite compact and
ideal for monolithic implementation. Besides, the circuit also exhibits high output impedance
making it suitable for current-mode cascading.
Another class of first order single voltage input filter circuits based on DVCC that
provide five independent outputs simultaneously; three of which are voltage outputs and three
as current outputs were proposed by Maheshwari [18]. Shown in Figure 9, the DVCC based
circuit takes a single voltage-mode input and is able to provide low-pass, high-pass and all-
pass filter functions both in the voltage- as well as current-mode.
Y1
Y2
Z+
Z-
DVCC
X
Y1
Y2 Z+
Z-
DVCC
XC
R
VIN
VLP
RL
VAP
IAP
ILPVHP
RL
IHP
Figure 9. DVCC based first order filter with single input and six outputs [18]
Linear and Non-Linear Applications of CMOS Dvcc
7
The various voltage-mode and current-mode filter transfer functions for the circuit [18]
are reproduced below.
(9)
(10)
(11)
[
] (12)
[
] (13)
[
] (14)
It is to be noted that six first order responses (9 – 14) are simultaneously obtained from
the circuit with a total active and passive component count of five, which makes the circuit
unique till date.
A voltage-mode all-pass section was also reported in [19] and is reproduced in Figure 10.
The canonical first-order all-pass section employed a single grounded capacitor and two
DVCCs, whose RX was tuned through bias voltage, thereby ensuring a resistor-less
realization. The idea of controlling the RX for DVCC was in fact first introduced in this work
[19]. This development proved to be a significant one for realizing active-C tunable networks
using DVCC.
Y1
Y2
Z+
Z-
DVCC
X
Y1
Y2
Z+
Z-
DVCC
X
VOUT
VIN
C
Figure 10. The first voltage-controlled voltage-mode first-order all-pass section using DVCC[19]
Sudhanshu Maheshwari, Mohd. Samar Ansari and Syed Atiqur Rahman
8
The transfer function of the circuit [19] was shown to be
(15)
As can be seen from (15), the circuit of Figure 10 indeed acts as a voltage-mode all-pass
filter and exhibits a gain of unity for all frequencies (ideally) and has a frequency dependent
phase characteristics given by
( ) ( ) (16)
D. Biquadratic Filters
Elwan and Soliman [7] proposed a CMOS continuous time filter for analog signal
processing in the current-mode domain. The electronically tunable band-pass filter, shown in
Figure 11 [7], was shown to have the following transfer function
(17)
Where
( ) (18)
( ) (19)
The center frequency and quality factor of the current-mode band-pass filter of Figure 11
were given as
√
(20)
√
(21)
Equations (20) and (21) suggest that the circuit does not exhibit non-interactive control
over ωo and Q, which is a serious drawback.
Next, the differential all-pass and notch filter circuit is discussed [20]. From Figure 12, it
can be seen that the circuit accepts differential voltage-mode input (Vid) and provides a
differential output (Vod). By selecting Rz1 = Rz2 = Rz and Cz1 = Cz2 = Cz, the differential-mode
gain of the circuit of Figure 12 [20] can be written as
( )
( ) (22)
Linear and Non-Linear Applications of CMOS Dvcc
9
Y1
Y2
Z+
Z-
DVCC
X Y1
Y2 Z+
Z-
DVCC
X
C1
VC1
IBP
VC2
C2
IIN
Figure 11. DVCC based Current-mode MOSFET-C Bandpass Filter [7]
Y1
Y2 Z+
Z-
DVCC
X
Vi1
Vi2
Vo1
Vo2
RX
CX
+
-
+
-Vid Vod
Rz1
Cz1
Cz2
Rz2
Figure 12. Differential Voltage-mode All-Pass and Notch Filter Circuit [20]
From (14), it is evident that voltage-mode all-pass filter function is realized by choosing
CX = 2CZ and RZ = 2RX. Similarly, the notch can be realized by selecting CX = CZ and RZ =
RX.
Chen and Shen proposed a three-input, five-output, universal capacitor-grounded voltage-
mode filter using DVCCs [21]. The circuit, shown in Figure 13, employs two differential
voltage current conveyors as active elements together with two grounded capacitors and four
resistors as passive elements. The proposed configuration can be used as either in the single-
input five-output or three-input two-output configuration. The circuit was shown to
simultaneously realize five different generic filtering signals: lowpass, bandpass, highpass,
bandreject, and allpass while enjoying the advantage of all grounded capacitors. Depending
on the status of the three biquad input voltages, Vi1, Vi2, and Vi3, different filter functions are
reproduced in the table below.
Table I. Obtaining different filter functions from the circuit of figure 13
Node Voltages Filter Function Vi1 Vi2 Vi3 Vout
Low-Pass Vin 0 0 Vo2
Band-Pass 0 Vin 0 Vo3
High-Pass 0 0 Vin Vo3
Band-Stop Vin 0 0 Vo3
All-Pass Vin Vin 0 Vo3
Sudhanshu Maheshwari, Mohd. Samar Ansari and Syed Atiqur Rahman
10
Y1
Y2
Z+
Z-
DVCC
X
Y1
Y2Z+
Z-
DVCC
X
Vo4
Vi2
R1
Vo5
R4
C1
Vi3
R3
C2
Vi1
Vo3
Vo2
Vo1
Figure 13. DVCC based voltage-mode biquad [21]
Y1
Y2
Z+
Z-
DVCC
X Y1
Y2 Z+
Z-
DVCC
X
Y1
Y2
Z+
Z-
DVCC
X
ILP
IBPIHP
IIN
R2 R3 R4
R1 C1 C2
Figure 14. Current-mode KHN biquad based on DVCC [15]
A multi-functional filter based on the Kerwin-Huelsman-Newcomb (KHN) topology and
utilizing DVCCs was also proposed by Ibrahim, Minaei and Kuntman [15]. The circuit,
shown in Figure 14, operates in the current-mode domain and employs three DVCCs as active
elements together with two capacitors and four resistors as passive elements, all of which are
grounded. The circuit simultaneously provides the three basic filter functions, viz. band-pass,
high-pass and low-pass functions. The notch and all-pass functions can be obtained by
connecting appropriate output currents directly without using additional active elements.
The circuit of Figure 14 was shown to yield the following current-mode transfer
functions
( )
(
)
(23)
( )
(
)
(24)
( )
(
)
(25)
Linear and Non-Linear Applications of CMOS Dvcc
11
The natural angular frequency ωo and quality factor of the filters can be expressed as
√
(26)
√
(27)
It can be seen that the three basic filter functions (HP, BP and LP) are obtained
simultaneously, all at high impedance outputs. Also, for R1=R2, a notch function can be
realized by connecting high-pass and low-pass outputs, i.e. by adding the corresponding
currents without using additional active elements. Moreover, for R1=R2=R3, an all-pass
function can be obtained by connecting high-pass, low-pass and band-pass outputs. Equations
(26) and (27) further show that non-interactive control of ωo and Q is only possible by
varying resistance ratio. Other realizations of the KHN biquad using three or four DVCCs
appear in [22].
Another universal biquadratic filter that needs to be mentioned is the one proposed by
Horng et al. [23]. The circuit, which is reproduced in Figure 15, employs three DVCCs and
six passive components to yield five voltage-mode filter functions viz. low-pass, high-pass,
band-pass, band-stop and all-pass simultaneously. An attractive feature of the circuit from
monolithic implementation point of view is that all the capacitors are grounded.
The transfer functions, resonance angular frequency and quality factor for the circuit of
Figure 15 are reproduced below.
(28)
(29)
(30)
(31)
(32)
√
(33)
√
(34)
Equations (33) and (34) show that ωo and Q can be tuned independent of each other,
thereby making the circuit useful.
Sudhanshu Maheshwari, Mohd. Samar Ansari and Syed Atiqur Rahman
12
Y2
Y1
DVCC
X
Z+
Y2
Y1
DVCC
X
Z+
Y2
Y1
DVCC
X
Z+
R2R1
C1 C2
Vo1 Vo2
Vo3
VIN
Vo5
R3 R4
Vo4
Figure 15. Voltage-mode Universal Biquadratic Filter employing DVCCs [23]
E. Sinusoidal Oscillators and Function Generators
Another class of DVCC-based analog circuit applications that will be considered includes
sinusoidal oscillators and function generators. The first such circuit is the four phase clock
generator [19]. As can be seen in Figure 16, an all-pass section is used to provide a 90-degree
phase shift to an input sinusoidal signal and then two DVCC based comparators are used to
obtain four distinct square waveforms in phase quadrature. The use of DVCC as comparator
was shown for the first time in [19], which led to further new applications, as would also be
shown later in this chapter.
Y1
Y2
Z+
Z-
DVCC
X
Y1
Y2
Z+
Z-
DVCC
X
V1
V2
VIN
APS
V3
V4
Figure 16. Four phase clock generator using an all-pass section [19]
Y2
Y1
DVCC
X
Z+
R3
R1
Z-
R2
C1
C2
Iout
Figure 17. Current-mode Single Resistance Controlled Oscillator using DVCC [24]
Linear and Non-Linear Applications of CMOS Dvcc
13
Y1
Y2
Z+
Z-
DVCC
X
Y1
Y2
Z+
Z-
DVCC
X
IOUT
Y4Y2 Y3Y1
VOUT
Figure 18. Single element controlled mixed-mode sinusoidal oscillator based on DVCC [25]
A DVCC-based circuit which belongs to the category of single element controlled
oscillators is discussed next. The circuit, attributed to Gupta and Senani [24] and shown in
Figure 17, employs only a single DVCC to provide a current-mode sinusoidal output which
can be controlled by varying a single resistance. It needs to be mentioned that the particular
circuit was the first single-resistance controlled current-mode oscillator using DVCC which
employed all grounded capacitors.
The frequency and condition of oscillation for the circuit are reproduced below.
FO:
√
(35)
CO: [
] (36)
Since R1 does not feature in FO above, the CO may be set by varying R1 and then FO can
independently be selected by choosing a suitable value of R2.
Another second order sinusoidal oscillator based on the DVCC is discussed next. By
proper selection of the admittances Y1 through Y4, eight different oscillator structures were
obtained from the circuit shown in Figure 18 [25]. For instance, by setting
a mixed-mode sinusoidal oscillator is obtained where the frequency and condition of
oscillation are given as
FO:
√
(37)
CO: (38)
An evident advantage of the above circuit is that the FO (frequency of oscillation) can be
adjusted by tuning a single element (capacitor or resistor) without disturbing the CO
(condition of oscillation) while the CO can be adjusted by tuning another element (resistor or
capacitor) without disturbing the FO thereby ensuring independent amplitude and frequency
control of oscillations.
Sudhanshu Maheshwari, Mohd. Samar Ansari and Syed Atiqur Rahman
14
Y1
Y2
Z+
Z+
DVCC
X
Y1
Y2
Z+
Z-
DVCC
X
R2
C2C3
R1
C1
R3
Figure 19. Voltage-mode fully-differential sinusoidal oscillator [26]
Next, a fully differential sinusoidal oscillator is discussed. Khan and Beg [26] proposed a
DVCC based oscillator capable of producing a fully differential sinusoidal waveform. A
voltage-mode fully-differential all-pass filter section is used to realize a voltage-mode fully-
differential sinusoidal quadrature oscillator by cascading it with a differential integrator as
shown in Figure 19. The use of one floating capacitor poses integration problems. Moreover,
the R1-C1 series combination connected at the X-terminal limits the frequency generated by
the circuit.
For the circuit of Figure 19, the frequency and condition of oscillation were found out to
be [26]
FO:
√
(39)
CO: (40)
Unlike the class of second order oscillators, a third order quadrature oscillator was
recently proposed as an application of a new first order topology by Maheshwari [13]. The
DVCC paper [13] was first of its kind, encompassing analog circuit applications ranging from
first-order to third-order. It is to be noted that Maheshwari [13, 27] was first to propose third
order quadrature oscillators based on DVCC. The circuit in [13] provided four quadrature
voltage outputs, whereas the circuit in [27] was shown to provide seven distinct voltage and
current outputs. The idea of voltage controlled DVCC [19] was further utilized in [27] to
realize a voltage controlled quadrature oscillator of third order.
III. PROPOSED CIRCUITS
This section presents some new applications of the DVCC. Some neural circuits based on
the DVCC are proposed which find applications in solving simultaneous linear equations,
linear programming problems and quadratic programming problems. Further, new
applications of the DVCC in realizing digital logic functions like AND/NAND and OR/NOR
are also explored.
Linear and Non-Linear Applications of CMOS Dvcc
15
A. Linear Equation Solver
The quest for methods to solve systems of linear equations has been going on for almost
two millennia [28-30]. The state-of-art linear equation solvers use software algorithms.
However, since most computer algorithms are sequential, the time to arrive at the solution is
governed by the problem size and the hardware on which the software is run. For real-time
systems, such a solution time of the order of milliseconds may not be appropriate. For
improved solution time, the inherent parallel processing of neural networks may be utilized
[30]. This section presents a neural circuit employing non-linear feedback to solve systems of
linear equations.
Let the set of simultaneous linear equations to be solved be
( ) (41)
( ) (42)
Where V1 and V2 are the two variables and a11, a12, a21, a22, b1 and b2 are constants.
Figure 20 shows the proposed hardware for solving a set of 2 simultaneous linear equations in
2 variables.
As can be seen from Figure 20, individual equations from the set of equations to be
solved are passed through non-linear synapses. CMOS Differential Voltage Current
Conveyors (DVCCs) are used as voltage-mode comparators and opamps are utilized to
emulate the operation of neurons. The DVCC is able to provide multiple current outputs
simultaneously, at the Z+ ports, which are scaled suitably to generate weighted currents at the
input of the neurons. For this purpose, use is made of the W/L ratios of the MOSFETs
employed in the output stage of the DVCC. Rp1 & Rp2 are the parasitic resistances and Cp1 &
Cp2 are the parasitic capacitance that present themselves at the inputs of the Opamps used in
the realization of neurons. These parasitic components are included to model the dynamic
nature of the Opamps. For the DVCC, we know that
( ) (43)
( ) (44)
For the DVCC, comparator action can be achieved by directly grounding the X-terminal.
The use of DVCC as a voltage-mode comparator was first given in [19]. For such a case, the
current in the X-port will saturate and can be written as
( ) ( ) (45)
By virtue of DVCC action, this current will be transferred to the Z+ ports. For DVCC-1,
the output currents at the Z+ ports can be written as
( ) ( ) (46)
Sudhanshu Maheshwari, Mohd. Samar Ansari and Syed Atiqur Rahman
16
( ) ( ) (47)
Where and are the scaling factors introduced during the current copying process
from the X-port to the Z-port. Ideally, these factors are unity. Similarly, for the second
comparator, we may write
( ) ( ) (48)
( ) ( ) (49)
Where is the open-loop gain of the comparator (practically very high) and is the
maximum (saturated) current at the output of the comparator. Writing node equation for the
nodes „A‟ and „B‟ in Figure 20, we get
( – ) ( – )
( )
( ) ( )
( )
Moreover, it is known that in the case of symmetric connections, the equations of motion
for this network of analog processing units corresponds to a stable system, in which the
output voltages of all amplifiers become constant [31]. Also, the stable states of the network
are the local minima of a quantity known as the computational energy „E‟ of the system (more
commonly referred to as „Lyapunov function‟ or „Energy Function‟) [32].
b1 x1
b2
V2Rp2 Cp2
B
V1Rp1 Cp1
As11IX1
Z1+
Z2+
X
X
Y2
Y1
Y2
Y1
DVCC-1
DVCC-2
a21V1 + a22V2
a11V1 + a12V2s21IX1
s12IX2
s22IX2
Z1+
Z2+
Figure 20. Proposed DVCC-based CMOS compatible circuit to solve a system of simultaneous linear
equations in 2 variables
Linear and Non-Linear Applications of CMOS Dvcc
17
For the network of Figure 20, it can be shown that the Energy Function „E‟ of is related
to the voltages V1 and V2 by
( ) ( ) [
] (52)
( ) ( ) [
] (53)
Also, if „E‟ is the Energy function, it must satisfy the following conditions for stability
[33].
(54)
And
(55)
Comparing equations (50) & (52), and (51) & (53) according to (54) and (55), we get
Although Figure 20 shows the proposed circuit as applicable to a system of simultaneous
linear equations in two variables, the technique can be readily extended to solve n-
simultaneous linear equations in n- variables. Moreover, for obtaining weighted (scaled)
current outputs use was made of the (W/L) ratios of the transistors comprising the Z+ stages
of the DVCCs used in Figure 20. This results in a particular circuit implementation „hard-
wired‟ to solve a given set of equations. Methods to scale the currents by a mechanism
external to the circuit need to be investigated to impart flexibility to the circuit.
B. Linear Programming Circuit
In a linear programming problem (LPP), a linear function is optimized subject to certain
linear constraints [34-36]. Linear programming can be applied to various fields of study. It is
used most extensively in business and economics, but can also be utilized for some
engineering problems. Industries that use linear programming models include transportation,
energy, telecommunications, and manufacturing. It has proved useful in modeling diverse
types of problems in planning, routing, scheduling, assignment, and design.
Let the function to be minimized be
(56)
Subject to
Sudhanshu Maheshwari, Mohd. Samar Ansari and Syed Atiqur Rahman
18
(57)
(58)
Where V1, and V2 are the variables and aij, bi and cij are constants. The proposed neural-
network based circuit to minimize (56) subject to (57, 58) is presented in Figure 21. As can be
seen from Figure 21, individual equations from the set of constraints are passed through non-
linear synapses which are realized using DVCC-based unipolar comparators. The outputs of
the comparators are fed to neurons having weighted inputs. These weighted neurons are
realized by using opamps where the scaled currents coming from various comparators act as
weights. The parasitic components that need to be included to model the dynamic nature of
the opamps are also considered (as was done for the previous circuit) and are shown in Figure
21.
Node equation for nodes „A‟ and „B‟ give the equations of motions of the first and second
neurons as
[
] [
] (59)
Where, and represent the currents and after passing through
diodes as shown in Figure 21 and
(60)
And,
[
] [
] (61)
Where, and represent the currents and after passing through
diodes as shown in Figure 21 and
(62)
In equations (59) and (61), ui is the internal state of the i-th neuron and is the current
scaling factor at the j-th output of i-th DVCC. As is shown later in this section, these weights
are governed by the constraint inequalities (57, 58). Using (46, 47) in (59) and (61) results in
(63) and (64) given below.
[ ( – ) ]
[ (
– ) ] [
] [
] (63)
Linear and Non-Linear Applications of CMOS Dvcc
19
s11IX1’
b1
x1
s11Ix1
Z+
1
XY2
Y1
DVCC-1
a11V1+ a12V2
Z+
2
s12Ix1
b2
x1s21Ix2
Z+
1
XY2
Y1
DVCC-2
a21V1+ a22V2
Z+
2
s22Ix2
+
_
V1
+_
Rp2C2V2
Rp1 C1
R1
R2
A
B
P1
P2
s21IX2’
s22IX2’
s12IX1’
Figure 21. Proposed neural network for solving a linear programming problem in 2 variables subject to
2 linear constraints
[ ( – ) ]
[ (
– ) ] [
] [
] (64)
Moreover, it can be shown that the network of Figure 21 may be associated with an
„energy‟ function of the form
( – )
( – )
( )
∫
∫
(65)
From (68), it follows that
( ) ( )
[
] (66)
Also, if „E‟ is the Energy Function, it must satisfy the following condition for stability
[33].
(67)
Where „K‟ is a constant of proportionality and has the dimensions of resistance.
Comparing (63) and (66) according to (67) yields
Sudhanshu Maheshwari, Mohd. Samar Ansari and Syed Atiqur Rahman
20
⁄
⁄ (68)
(69)
Analysis on similar lines can be performed to obtain the values of the synaptic weights
for the second neuron and are presented in (70) and (71) below.
⁄
⁄ (70)
(71)
The values of the current scaling factors (sij) of Figure 21 can be calculated by choosing a
suitable value of „K‟ and then using (68, 70).
C. Quadratic Programming Circuit
Quadratic Programming problems are very important in the field of optimization [37-39].
They arise in many applications such as constrained least mean square estimation and the
classical newsvendor problem. Besides its wide applications, quadratic programming is also
forms a basis for solving some general nonlinear programming problems as most higher order
problems can usually be approximated by a second order function. Traditional methods for
solving quadratic programming problems typically involve an iterative process, but long
computational time limits their usage. There is an alternative approach to solution of this
problem. It is to exploit the artificial neural networks which can be considered as an analog
computer relying on a highly simplified model of neurons [31].
Let the quadratic function to be minimized be
(72)
Subject to
(73)
(74)
Where V1, and V2 are the variables and aij, bi and cij are constants. The proposed neural-
network based circuit to minimize (72) subject to (73, 74) is presented in Figure 22.
As was done in the case of linear programming circuit, individual equations from the set
of constraints are passed through non-linear synapses which are realized using DVCC-based
unipolar comparators. The outputs of the comparators are fed to neurons having weighted
inputs. These weighted neurons are realized by using opamps where the scaled currents
coming from various comparators act as weights. Weighing is also provided by the self-
Linear and Non-Linear Applications of CMOS Dvcc
21
feedback resistances (R11 & R22 for neurons 1 & 2 respectively) and the cross-feedback
resistances (R21 and R12).
Node equation for nodes „A‟ and „B‟ give the equations of motions of the first and second
neurons as
[
] [
] (75)
Where, and represent the currents and after passing through
diodes as shown in Figure 22 and
(76)
And,
[
] [
] (77)
Where, and represent the currents and after passing through
diodes as shown in Figure 22 and
(78)
b1
x1s11Ix
Z+
1
XY2
Y1
DVCC-1
a11V1+ a12V2
Z+
2
s12Ix
b2
x1s21Ix
Z+
1
XY2
Y1
DVCC-2
a21V1+ a22V2
Z+
2
s22Ix
+
_V1
+_
Rp2Cp2
V2
Rp1 Cp1
R11
R22
R21
R12
A
B
s11IX’
s12IX’
s21IX’
s22IX’
Figure 22. Proposed circuit to solve quadratic programming problem in 2-variables with 2 linear
constraints
Sudhanshu Maheshwari, Mohd. Samar Ansari and Syed Atiqur Rahman
22
In equations (78) and (80) ui is the internal state of the i-th neuron and is the current
scaling factor at the j-th output of i-th DVCC. As is shown later in this section, these weights
are governed by the constraint inequalities (73, 74). Using (46, 47) in (75) and (77) results in
(79) and (80) given below.
[ ( – ) ]
[ (
– ) ] [
] [
] (79)
[ ( – ) ]
[ (
– ) ] [
] [
] (80)
Moreover, it can be shown that the network in Figure 22 can be associated with an
Energy Function „E‟ of the form
( – )
( – )
( )
∫
∫
(81)
From (81), it follows that
( )
(
) [
] (82)
Also, if „E‟ is the Energy Function, it must satisfy the following condition for stability
[33].
(83)
Where „K‟ is a constant of proportionality and has the dimensions of resistance.
Comparing (79) and (82) according to (83) yields
(84)
(85)
The values of the current scaling factors (sij) of Figure 22 can be easily calculated by
setting the value of „K‟ to be equal to ⁄ and then using (84). The values of the weight
resistances connected to the first neuron, R11 and R21, can be obtained by using (85). Analysis
Linear and Non-Linear Applications of CMOS Dvcc
23
on similar lines can be performed to obtain the values of the synaptic weights for the second
neuron.
D. Implementation of Logic Gates Using DVCC
Another application area where the DVCC can be put to use is Digital Logic Gates.
Although implementations of various digital logic gates are abundant in various technologies
including CMOS, the use of analog building blocks like the DVCC has never been considered
before. The circuit for 2-input AND/NAND gate, presented in Figure 23, is a first attempt to
showcase the full potential of the DVCC and is by no means a comparison with the existing
implementations, which employ fewer transistors.
The working of the AND gate of Figure 23 can be understood as follows. As long as V2
is low, transistor M1 remains off leading to IX = 0 (irrespective of the voltage applied at Y1
terminal) which causes IZ and consequently VAND to be zero. When V2 becomes high,
transistor M1 conducts and the current IX will be governed by the presence or absence of V1.
Under such a condition, for the case of V1 being low (zero), IX will be zero and consequently,
VAND is low. For the case when V1 is high, the output voltage VAND will be high. The input
and output for the AND gate are related as
( )
Similarly, the circuit for OR/NOR gates can be obtained if the Y3 terminal of the DVCC
is also utilized [13]. An analysis similar to that done for the AND/NAND circuit shows that
the circuit of Figure 24 does indeed works as an OR/NOR gate. It may be mentioned that
VOR/VNOR may be needed to be passed through an additional comparator if a binary digital
output is required.
Y2
Y1
DVCC
X
Z+V1
V2
R1
VAND
M1
R1
VNANDZ-
Figure 23. AND/NAND gate realization using DVCC
Figure 24. OR/NOR gate realization using DVCC
Y2
Y1
DVCC
X
Z+V1
V2
R1
VOR
M1
R1
VNOR
M2V1
Y3 Z-
Sudhanshu Maheshwari, Mohd. Samar Ansari and Syed Atiqur Rahman
24
IV. SIMULATION RESULTS
This section deals with the simulation results for the proposed circuits. Simulations were
carried out using PSPICE program. The circuit for DVCC shown in Figure 4 was utilized and
standard 0.5 micron CMOS parameters were used for simulation purposes. For the opamp,
use was made of the LMC7101A CMOS opamp from National Semiconductor. The sub-
circuit file for this opamp is available in Orcad Model Library.
The proposed circuit for solving linear equations was tested for solving the following set
of 2 simultaneous linear equations in 2 variables.
[
] [ ] [
] ( )
For the above set of equations, the required weighted currents that need to be generated
are given as: s11 = 2, s12 = 2, s21 = 1 and s22 = 1. Scaling of the Z-port currents was done by
varying the aspect ratios of the transistors in the output stage of the DVCC. The circuit used
for PSPICE simulation is presented in Figure 25. As can be seen, some additional circuitry is
required to generate the proper input voltages of the comparators [30]. The values of the
different resistances shown in Figure 23 are: Re11 = 2K; Re12 = 4K; Re13 = 4K; Re21 = 3K; Re22
= 3K; Re23 = 3K.
PSPICE simulations were performed for the chosen set of equations and the results are
given in Figure 26. As can be verified from the figure, the simulation results are found to
agree exactly with the algebraic solution.
Figure 25. The proposed circuit applied to solve 2 linear equations in 2 variables
b1 x1
b2
Re21
Re22
Re23
V2
Re11
Re12
Re13
V1
Z+
Z+
X
X
Z+
Z+
Y2
Y1
Y2
Y1
DVCC-1
DVCC-2
Linear and Non-Linear Applications of CMOS Dvcc
25
Figure 26. PSPICE simulation results of the circuit to solve (86)
Next, the proposed network for solving linear programming problems was tested using
PSPICE simulation. The circuit was used to minimize the objective function
(87)
subject to
The values of resistances acting as the weights on the neurons are obtained from (69, 71).
Towards that end, first the value of was computed for the DVCC which was found to be
1.091 milli-mhos. Choosing ⁄ allows us to fix the current scaling factors as given
below:
Routine mathematical analysis of (87, 88) yields: = 0, = 1. The results of PSPICE
simulation are presented in Figure 27. From the plots of the neuron output voltages, it can be
seen that V(1) = 0 V and V(2) = 1 V which correspond exactly to the algebraic solution
thereby confirming the validity of the approach.
(88)
Sudhanshu Maheshwari, Mohd. Samar Ansari and Syed Atiqur Rahman
26
Figure 27. Simulation results for the proposed circuit for solving linear programming problems (87, 88)
The proposed neural circuit for solving quadratic programming problems was tested by
minimizing the objective function
(89)
subject to
The values of resistances acting as the weights on the neurons are obtained from (85).
The corresponding values of the weight resistances are found to be: R11 =305.53Ω; R21
=229.14Ω; R12 =229.14Ω; R22 =183.318Ω. The values of the current scaling factors were
found out to be: s11 = 1; s12 = -1; s21 = 1 ; s22 = 1.
Routine mathematical analysis of (89, 90) yields: = -0.584, = 0.416. The results of
PSPICE simulation are presented in Figure 28. From the plots of the neuron output voltages,
it can be seen that V(1) = -0.58 V and V(2) = 0.41 V which are very near to the algebraic
solution thereby confirming the validity of the approach.
Next, the AND gate of Figure 23 was verified. The results of simulation are presented in
Figure 29 where the voltages V1 and V2 are the inputs to the AND gate and voltage V4 is the
AND-ed output.
(90)
Linear and Non-Linear Applications of CMOS Dvcc
27
Figure 28. Simulation results for the chosen 2 – variable problem (89, 90)
Figure 29. Simulation results for the proposed AND gate using DVCC
An important point that needs to be discussed is the VLSI implementability issue of the
existing and proposed CMOS circuits. It is well known that in standard CMOS technology,
Sudhanshu Maheshwari, Mohd. Samar Ansari and Syed Atiqur Rahman
28
processing variations prohibit the fabrication of individual resistances with precise values. It
is, however, possible to fabricate a pair of resistors having the same value, although the exact
value may not necessarily be predicted. Considering this fact, one way of implementing
resistances in CMOS analog circuits is to tweak the design such that the critical (performance
deciding) factors are governed by resistance-ratios [40]. Wherever this is not possible, use
may be made of MOS-based resistors. The function of a resistor may be obtained using
MOSFETs to realize voltage-controlled transconductances or by linearizing the transistor‟s
region of operation in the triode region [41].
V. CONCLUSION
This chapter was based on circuit applications of differential voltage current conveyor in
CMOS technology. A survey of existing linear and non-linear circuits was presented. Some
new applications of DVCC were also proposed. The chapter explores a wide variety of analog
and a few digital applications and is expected to both provide a deep insight into the subject
and also further enhance the existing knowledge.
REFERENCES
[1] Smith, K. C. & Sedra, A. S. (1968). “The current-conveyor-a new circuit building
block,” IEEE Proc., Vol. 56, Pages 1368-69.
[2] Maheshwari, S. & Khan, I. A. (2005). “Current controlled third order quadrature
oscillator,” IET Circuits Devices Syst., Vol. 152, No. 6, Pages 605-607.
[3] Khan, I. A. & Ahmed, M. T. (1986). “Electronically tunable first-order OTA-capacitor
filter sections,” Int. J. of Electronics, Vol. 61, No. 2, Pages 233-237, August.
[4] Biolek, D. (2003). “CDTA–Building block for current-mode analog signal processing”,
Proc. ECCTD Poland, Pages 397-400.
[5] Minaei, S. & Yuce, E. (2008). “A new full-wave rectifier circuit employing single dual-
X current conveyor,” Int. J. of Electronics, Vol. 95, No. 8, Pages 777-784, August.
[6] Pal, K. (1989). “Modified current conveyors and their applications,” Microelectronics
Journal, Vol. 20, No. 4, Pages 37-40.
[7] Elwan, H. O. & Soliman, A. M. (1997). “Novel CMOS differential voltage current
conveyor and its applications,” IEE Proc. Circuits Devices Syst., Vol. 144, No. 3, Pages
195-200, Jun.
[8] Moradzadeh, H. & Azhari, S. J. (2008). “High performance low-voltage QFG-based
DVCC and a novel fully differential SC integrator based on it,” IEICE Electronics
Express, Vol 5, No. 23, Pages 1017-1023, Dec.
[9] Mahmoud, S. A. (2008). “Low Voltage Wide Range CMOS Differential Voltage
Current Conveyor and its Applications,” Contemporary Engineering Sciences, Vol. 1,
No. 3, Pages 105-126.
[10] Minaei, S. & Ibrahim, M. A. (2005). “General configuration for realizing current-mode
first-order all-pass filter using DVCC,” Int. J. of Electronics, Vol. 92, No. 6, Pages 347-
356.
Linear and Non-Linear Applications of CMOS Dvcc
29
[11] Hu, J., Xu, T., Zhang, W. & Xia, Y. (2005). “A CMOS Rail-to-Rail Differential
Voltage Current Conveyor and its Applications,” IEEE Proc. ICCCAS, Vol. 2, Pages
1079-1083.
[12] Siripruchyanan, M. & Jaikla, W. (2007). “Floating Capacitance Multiplier Using
DVCC and CCCIIs,” Proc. Int. Symp. on Comm. and Info. Tech. (ISCIT '07).
[13] Hassan, T. M. & Mahmoud, S. A. (2010). “New CMOS DVCC realization and
applications to instrumentation amplifier and active-RC filters,” Int. J. Electronics and
Communications, Vol. 64, Pages 47-55.
[14] Maheshwari, S. (2009). “Analogue signal processing applications using a new circuit
topology,” IET Circuits Devices Syst., Vol.3, No.3, Pages 106-115.
[15] Ibrahim, M. A., Minaei, S. & Kuntman, H. (2005). “A 22.5 MHz current-mode KHN-
biquad using differential voltage current conveyor and grounded passive elements,” Int.
J. Electronics and Communications, Vol. 59, No. 5, Pages 311-318, Jul.
[16] Minaei, S. & Temizyurek, C. (2003). “Dual Input All-Pass Filter using DVCC,” Proc.
Int. Symposium on Signals, Circuits and Systems (SCS), Vol. 2, Pages 477-480.
[17] Maheshwari, S. (2007). “Novel cascadable current-mode first order all-pass sections,”
Int. J. Electronics, Vol. 94, No. 11, Pages 995-1003, Nov.
[18] Maheshwari, S. (2008). “CMOS Compatible Single Input Five/Six Outputs First Order
Filter Circuits,” J. Active Passive Elec. Dev., Vol. 3, Pages 125-133.
[19] Maheshwari, S. (2008). “A Canonical Voltage-Controlled VM-APS with a Grounded
Capacitor,” Circuits, Systems, and Signal Processing, Vol. 27, No. 1, Pages 123-132,
Feb.
[20] Ibrahim, M. A., Minaei, S. & Kuntman, H. (2006). “DVCC based differential-mode all-
pass and notch filters with high CMRR,” Int. J. Electronics, Vol. 9, No. 4, Pages 231-
240, Apr.
[21] Chen, H. P. & Shen, S. S. (2007). “A Versatile Universal Capacitor-Grounded Voltage-
Mode Filter Using DVCCs,” ETRI Journal, Vol. 29, No. 4, Pages 470-476, Aug.
[22] Soliman, A. M. (2008). “Generation and classification of Kerwin–Huelsman–Newcomb
circuits using the DVCC,” Int. J. Circ. Theor. and Appl., Vol. 37, Pages 835-855, Jun.
[23] Jiun-Wei Horng et al, (2006). “High Input Impedance Voltage-Mode Universal
Biquadratic Filter with One Input and Five Outputs Using Current Conveyors,”
Circuits, Systems, Signal Processing, Vol. 25, No. 6, Pages 767-777.
[24] Gupta, S. S. & Senani, R. (2000). “Grounded-Capacitor Current-mode SCRO: Novel
application of DVCC,” Electronics Letters, Vol. 36, No. 3, Pages 195-196.
[25] Kumar, P., Keskin, A. U. & Pal, K. (2007). “DVCC-Based Single Element Controlled
Oscillators Using All-Grounded Components and Simultaneous Current-Voltage Mode
Outputs,” Frequenz, Vol. 61, Pages 141-144.
[26] Khan, I. A. & Beg, P. (2009). “Fully Differential Sinusoidal Quadrature Oscillator
Using CMOS DVCC,” Proc. Int. Conf. On Communication, Computer and Power
(ICCCP'09), Muscat, Pages 196-198, Feb.
[27] Maheshwari, S. (2009). “Quadrature oscillator using grounded components with current
and voltage outputs,” IET Circuits Devices Syst., Vol.3, No.4, Pages 153-160.
[28] Gustafson, J. L. (2006). “The Quest for Linear Equation Solvers and the Invention of
Electronic Digital Computing,” Proc. JVA’06, IEEE John Vincent Atanasoff 2006
International Symposium on Modern Computing.
Sudhanshu Maheshwari, Mohd. Samar Ansari and Syed Atiqur Rahman
30
[29] Xia, Y., Wang, J. & Hung, D. L. (1999). “Recurrent Neural Networks for Solving
Linear Inequalities and Equations,” IEEE Trans. Circuits and Systems–1: Fundamental
Theory and Applications, Vol. 46, No. 4, Pages 452-462, April.
[30] Ansari, M. S. & Rahman, S. A. (2009). “A Novel Current-mode Non-Linear Feedback
Neural Circuit for Solving Linear Equations,” Proc. IMPACT – 2009, Aligarh, India,
Pages 284 – 287, Available on IEEE Xplore, DOI: 10.1109/MSPCT.2009.5164189
[31] Hopfield, J. J. & Tank, D. W. (1985). “„Neural‟ computation of decisions optimization
problems,” Biological Cybernetics, Vol. 52, Pages 141-152.
[32] Tank, D. W. & Hopfield, J. J. (1986). "Simple neural optimization networks: an A/D
converter, signal decision network, and linear programming circuit," IEEE Trans. Circ.
Syst, Vol. 33, Pages 533-541.
[33] Rahman, S. A. Jayadeva, & Dutta Roy, S. C. (1999). "Neural network approach to
graph colouring," Electronics Letters, Vol. 35, Pages 1173-1175.
[34] Wang, J. (1993). “Analysis and Design of a Recurrent Neural Network for Linear
Programming”, IEEE Trans. Circuits and Systems–1: Fundamental Theory and
Applications, Vol. 40, No. 9, Pages 613-618, Sept.
[35] Xia, Y. (1995). “A New Neural Network for Solving Linear Programming Problems
and its Application,” IEEE Trans. Neural Networks, Vol.7, No.2, Pages 94-103, March.
[36] Malek, A. & Yari, A. (2005). “Primal–dual solution for the linear programming
problems using neural networks,” Applied Mathematics and Computation, Vol. 167,
Pages 198-211.
[37] Kennedy, M. P. & Chua, L. O. (1988). “Neural networks for nonlinear programming,”
IEEE Trans. Circ. Syst., Vol. 35, Pages 554-562.
[38] Wu, X. Y., Xia, Y. S., Li, J. & Chen, W. K. (1996). “A high performance neural
network for solving linear and quadratic programming problems,” IEEE Trans. Neural
Networks, Vol. 7, No. 3, Pages 643-651.
[39] Malek, A. & Alipour, M. (2007). “Numerical solution for linear and quadratic
programming problems using a recurrent neural network,” Appl. Math. Computation,
Vol. 192, Pages 27-39.
[40] Uyemura, J. P. (2006). “Chip Design for Submicron VLSI: CMOS Layout and
Simulation,” Cengage Learning, First Ed., Pages 381-382.
[41] Mahmoud, S. A., Elwan, H. O. & Soliman, A. M. (1997). “Generation of CMOS
Voltage-Controlled Floating Resistor,” Microelectronics Journal, Vol. 28, Pages 627-
640.
THE PROCEEDINGS OF IEEE SPONSORED INTERNATIONAL CONFERENCE RASIET-07
IET ALWAR (RAJ) ON 22ND
-23RD
DEC 07/VOL.2
570
A Non – Linear Feedback Neural Network for Graph
Coloring
S. J. Arif* S. A. Rahman* M. S. Ansari*
Abstract
This paper presents a hardware approach to the problem of graph coloring. In the graph coloring
problem, the task is to color, or label, the vertices such that no two adjacent vertices get the same
color. The neural network model presented also minimizes the number of colors used. This
additional capability provides for a higher efficiency of resource usage as well as having
application to other problems that map to graph coloring. Hardware and simulation results on
random and benchmark problems have been presented. Test results are compared with previous
neural network techniques for graph coloring, and it is shown that this neural network model
provides a significant reduction in the number of colors required while maintaining a high
convergence rate.
1. Introduction
The map-coloring problem, famous for the four-color theorem, is a special case of the graph-coloring problem (GCP). Graph coloring has a large number of applications such as register allocation in digital computers [1], frequency channel assignment in mobile communication [2], and layer assignment in VLSI design. In GCP, values or colors are assigned to different nodes of a graph with a condition that no two adjacent nodes of the graph should have the same color, while the number of colors should also be minimized. For proper graph coloring, the adjacency constraints should be strictly satisfied. The Hopfield Neural Network and its modifications [3,4] have been applied to solve GCP, but the implementation requires N2
neurons and at most N4 weights for solving a graph of N nodes. This leads to significant consequences in terms of the feasibility of *Department of Electronics Engineering, Aligarh Muslim University, Aligarh
HNN realization in hardware. Another feedback Neural network architecture, which uses non-linear synapses, and has a transcendental Energy function, requires N neurons and at most N2 interconnections for solving a problem with N nodes [6]. In this paper, the neural circuit presented in [6] has been modified, in order to improve the solution quality. In the proposed network, the Energy function given in [6] has been modified by eliminating those local minima that correspond to large number of colors. This is achieved by replacing the bipolar comparators of the earlier circuit with unipolar comparators. It is seen that the number of colors is reduced.
2. Proposed Network For Graph
Coloring
Fig. 1 shows the ith neuron of proposed network. In the proposed circuit, output voltages of different neurons represent the colors of different nodes. Ci and ri denotes the internal capacitance and resistance of the ith neuron respectively, u
i is the internal state and Rii is the
feedback resistance of i-th neuron. The output
THE PROCEEDINGS OF IEEE SPONSORED INTERNATIONAL CONFERENCE RASIET-07
IET ALWAR (RAJ) ON 22ND
-23RD
DEC 07/VOL.2
571
of other neurons Vj (j = 1,2…N) are connected to the input of ith neuron through unipolar comparators. These unipolar comparators are
realized by connecting the diodes Dij in series with the bipolar comparators CMij. (j = 1,2…N).
ri
Ci
RC/D
xi1
xij
xiN
Vi
V1
Vj
VN
Ri1
Rij
RiN
Di1
Dij
DiN
CMi1
CMij
CMiN
Fig. 1 The proposed neural network based circuit for
graph coloring
If ith
neuron is connected to jth
neuron (aij =1),
then Vj is connected to the internal state of the
ith
neuron ui through comparator CMij and a
resistance Rij. This is a constant given by
Rij = aij Rc
where
1, if ith
node is connected by an
aij = edge to j
th node
0, otherwise
The transfer characteristic of the unipolar
comparator CMij is shown in fig. 2.
Fig. 2 Transfer Characterisitcs of CMij
The output of unipolar comparator CMij is
denoted by Xij . The transfer characteristic of
this comparator can be modeled by the
following equation.
( ) 2
tanh mijm
ij
VVVVX
−−=
β (1)
( )[ ] 1tanh2
−−= ijm
ij VVV
X β (2)
Where β is the gain of comparator which is
very high (β→∞), Vm is the biasing voltage of
comparator.
The equation of motion of ith
neuron is given as
∑≠=
−+
−=+
N
ijj ij
iij
ii
ii
i
iii
R
uX
R
uV
r
u
dt
duC
,1
(3)
which transforms to
∑∑≠=≠=
+−−−=N
ijj ij
ijN
ijj ij
i
i
i
ii
i
ii
iii
R
X
R
u
r
u
R
u
R
V
dt
duC
,1,1
(4)
THE PROCEEDINGS OF IEEE SPONSORED INTERNATIONAL CONFERENCE RASIET-07
IET ALWAR (RAJ) ON 22ND
-23RD
DEC 07/VOL.2
572
and thus ∑≠=
+−=N
ijj ij
ij
i
i
ii
iii
R
X
R
u
R
V
dt
duC
,1
(5)
where [ ]∑≠=
++=N
ijj ijiiii RrRR ,1
1111
Therefore, Ri is equivalent to parallel
combination of all the resistances connected to
the node at internal state ui.
Similarly, we have
ij
N
ijj
ij
i
iiii
ii XW
R
uVW
dt
duC ∑
≠=
+−=,1
(6)
Where ii
iiR
W1
= , ij
ijR
W1
=
The output of ith
neuron is related to the
internal state by
( )imi uVV λtanh−=
Where λ is the gain of the amplifier, which is
very high (λ→∞)?
The given circuit may be associated with the
following Lyapunov Function.
( )[ ]2
coshln4
11
,10
2
1
imij
N
ijj
mi
V
i
i
i
N
i
iic
VVVV
VdVu
RVWE
i
−−−−= ∑∫∑≠==
ββ
(7)
The second term in equation (1) is usually
neglected for λ→∞, and thus equation (7) can
be simplified to
( )[ ]2
coshln1
4 ,1
2
1
imij
N
ijj
cijm
i
N
i
iic
VVVVWa
VVWE −−−= ∑∑
≠==
ββ
(8)
The first term on the right hand side of (8) is
quadratic which tries to minimize the number
of colors by lowering the color values. The
second term has got a negative sign. Therefore,
the energy function Ec will be minimized if
second term is maximized. This happens when
the voltages corresponding to connected nodes
in a graph are far away from each other. The
first two terms on the right hand side are
balancing each other to color a graph properly.
The last term also contributes to lowering of
number of different colors by eliminating all
those local minima in the energy function for
which node voltages were negative.
3. Application Of The Proposed
Network To Graph Coloring
Problems
The proposed network was tested for various
example graphs as well as a standard
benchmark problem for graph coloring,
myciel3.col [5]. Both hardware and software
simulation tests were undertaken. As can be
seen from table 1,the proposed network gives a
solution to al the problems tested and in all the
cases the solution is very near the chromatic
number of the graph. As an example,
simulation runs of the proposed network for
the myciel3.col benchmarking problem
returned the best solution of 5 colors which is
very near to the chromatic number for
myciel3.col, that being 4.
The performance of the proposed network was
also compared with a network proposed earlier
[6]. Table 2 gives this performance
comparison. It is evident from the table that
substantial improvements have been achieved
both in the best and the average solutions.
Table 2. Performance comparison of the proposed
and existing networks
THE PROCEEDINGS OF IEEE SPONSORED INTERNATIONAL CONFERENCE RASIET-07
IET ALWAR (RAJ) ON 22ND
-23RD
DEC 07/VOL.2
573
4. Conclusion
A novel neural network based architecture to
solve graph coloring problems is presented.
The circuit was realized and simulated using
SPICE for an example set of eight different
graphs including the myciel3.col benchmark
problem. Simulation tests for the proposed
circuit ascertain its superiority over its
hardware predecessors. It is observed that both
the minimum and the average number of colors
are improved for the proposed network
References [1] T. Zeitlhofer and B. Wess, A Comparison of Graph
Coloring Heuristics for Register Allocation based on
Coelescing in Interval Graphs, Proc. ISCAS 04, vol. IV,
2004.
[2] N. A. El-Fishawy,M. M. Hadhood, S. Elnoubi, and
W. El-Sersy, A modified Hopfield neural network
algorithm for cellular radio channel assignment, Proc.
TENCON 2000, vol. 2, Sept 2000.
[3] K. Jin’no, H. Taguchi, T. Yamamoto, H. Hirose,
Dynamical hysteresis neural networks for graph
coloring problem, Proceedings of ISCAS03, vol. 5, 2003.
[4] W-D. Sun, H. Tamura, Z. Tang, M. Ishii, A Method
to Solve the Four-Coloring problem by the Neural
Network with Self-Feedback, Proc. SICE Annual
Conference in Fukui, Japan, Aug, 2003.
[5] J. Lui, W. Zhong, and L. Jiao, Comments on “The
1993 DIMACS Graph Coloring Challenge” and
“Energy Function-Based Approaches to Graph
Coloring”, IEEE Trans. On Neural Networks, vol. 17,
no. 2, Mar 2006.
[6] S. A. Rahman, Jayadeva and S. C. Dutta Roy, Neural
network approach to graph colouring, Electronics
Letters, vol. 35, Jul 1999
[7] Web resource for graph coloring benchmarks,
http://mat.gsia.cmu.edu/COLOR/instances.html
Mohd. Samar Ansari received the
B.Tech. degree in Electronics
engineering from the Aligarh Muslim
University, Aligarh, India, in 2001 and
the M. Tech. degree in Electronics
engineering, with specialization in
Electronic Circuit and System Design, in 2007. He is
currently a Lecturer in the Department of Electronics
Engineering, Aligarh Muslim University, where he
teaches Electronic Devices & Circuits and
Microelectronics. His research interests include
microelectronics, linear integrated circuit applications,
neural networks, and semiconductor physics.
Syed Atiqur Rahman did B.Sc.(Engg.) in Electrical
Engineering in 1988, and M.Sc.(Engg.) in Electronics
and Communication Engineering in 1994, from Aligarh
Muslim University, Aligarh. He obtained his PhD from
Electrical Engineering Department at IIT (Delhi) in
2007. He is currently a Reader in the Department of
Electronics Engineering, Aligarh Muslim University,
where he teaches neural networks and signals &
systems. His fields of interest are Electronic Circuits and
Artificial Neural Networks.
Syed Javed Arif did B.Sc.(Engg.) in Electrical
Engineering in 1988, and M.Sc.(Engg.) in 1990, from
Aligarh Muslim University, Aligarh. He is presently
working towards his PhD from Electronics Engineering
Department, Aligarh Muslim University, Aligarh. He is
currently a Lecturer in the Department of Electronics
Engineering, Aligarh Muslim University, where he
teaches Digital Electronics and Electronics Devices. His
fields of interest are Electronic Circuits and Artificial
Neural Networks.
THE PROCEEDINGS OF IEEE SPONSORED INTERNATIONAL CONFERENCE RASIET-07
IET ALWAR (RAJ) ON 22 -23 DEC 07/VOL.2
Solution of Simultaneous Linear Equations using Feedback
Neural Network
S. A. Rahman* M. S. Ansari*
Abstract Solving a linear system of equations is one of the most important computing problems. Existing
methods include the use of parallel computing architectures and complex software algorithms.
The time taken by the existing methods to solve a linear system of equations is not always short
and increases significantly with the number of variables. This paper presents a feedback neural
network based hardware solution to the problem of solving simultaneous linear equations. A
direct consequence of the use of neural networks is massively parallel processing and fast
convergence. This results in a faster and more compact hardware. Simulation results are
presented for a chosen set of equations and are found to agree exactly with the algebraic
solution.
I. INTRODUCTION
The solution of linear system of equations
continues to play an important role in
scientific computing. The problems to be
solved often are of very large size, so that
large computer resources or massive parallel
computer systems with distributed memory
are needed to solve them. This problem is
key to many real-time engineering
applications [1-6]. Examples are, among
others, real-time speech coding, image
processing, stochastic modeling, and
computer-aided realistic three-dimensional
image synthesis. When the problem at hand
is large, we must obtain its solution within a
reasonably short period of time. Existing
methods to solve a set of linear equations
also include Gaussian elimination and the
Cholesky decomposition.
*Department of Electronics Engineering,
Aligarh Muslim University, Aligarh
To determine approximate solutions to
general systems numerically on a computer,
the n-dimensional Newton's method may be
used.
Among the concurrent multiprocessor
computers the contemporary, relatively less
expensive systolic/wavefront arrays like
Kung’s are promising parallel architectures
[7]. These structures provide the necessary
computational power to tackle large
problems efficiently in terms of time. The
Block Data Parallel Architecture (BDPA) is
an even less expensive alternative than the
systolic/wavefront arrays for solving these
computationally intensive problems [8]. The
processing of data by blocks and the use of
commercially available processors makes
this architecture not only less expensive, but
also flexible and reconfigurable. The
architecture can be used, not only for
elementary DSP algorithms, but also for
complex algorithms usually associated with
linear algebra. However, even with the
extensive simulation tools developed for the
BDPA, there is still a fundamental need to
show the prospective power of the
architecture.
THE PROCEEDINGS OF IEEE SPONSORED INTERNATIONAL CONFERENCE RASIET-07
IET ALWAR (RAJ) ON 22 -23 DEC 07/VOL.2
This paper presents a hardware solution to
the problem of solving simultaneous linear
equations. The approach used is to employ
neural networks on the basis of an analog
circuit [2,4,6]. The most important
advantages of neural networks are massively
parallel processing and fast convergence.
Hence, architectures based on neural
networks have many computational
advantages over the traditional algorithms
running over a parallel computing system.
The following section lays the basic problem
and the mathematical formulation on which
the development of the proposed network
will be based. Section III contains the circuit
implementation of the proposed network for
a set of sample linear equations in two
variables. SPICE simulation results of the
proposed circuit are also presented. Some
concluding remarks are presented in Section
IV.
II. PROPOSED NEURAL NETWORK
BASED CIRCUIT
Let the simultaneous linear equations to be
solved be
𝑎11𝑉1 + 𝑎12𝑉2 − 𝑏1 = 0
𝑎21𝑉1 + 𝑎22𝑉2 − 𝑏2 = 0
Where V1 and V2 are the two variables and
a11, a12, a21, a22, b1 and b2 are constants.
The proposed neural-network based circuit
to solve the system of equations given above
is presented in fig. 1.
As can be seen from fig. 1, individual
equations from the set of equations to be
solved are passed through a non-linear
synapse. This non-linear synapse is in effect
a comparator which is realized using
Opamps. The outputs of the non-linear
synapses are fed to weighted neurons. These
weighted neurons are again realized by
using Opamps where the resistors R11
through R22 act as weights. Rpi and Cpi are
the parasitic resistance and capacitance that
present themselves at the input to the Opamp
used in the realization of the i-th neuron.
These parasitic components are included to
model the dynamic nature of the Opamp.
R11
R12
R21
R22
V1
V2
a11
V1+a
12V
2
b1
a21
V1+a
22V
2
b2
x1
u1
x2
u2
Rp1
Cp1
Rp2
Cp2
A
B
Fig. 1 The proposed Neural Network based circuit to solve a system of simultaneous linear equations in 2 variables.
Non-Linear Synapses Weighted Neurons
THE PROCEEDINGS OF IEEE SPONSORED INTERNATIONAL CONFERENCE RASIET-07
IET ALWAR (RAJ) ON 22 -23 DEC 07/VOL.2
Non-linear synapses are realized using
comparators. The comparator outputs can be
modelled by equations (1) and (2) given
below.
𝑥1 = 𝑉𝑚 𝑡𝑎𝑛ℎ 𝛽 𝑎11𝑉1 + 𝑎12𝑉2 − 𝑏1 (1)
𝑥2 = 𝑉𝑚 𝑡𝑎𝑛ℎ 𝛽(𝑎21𝑉1 + 𝑎22𝑉2 − 𝑏2) (2)
Where β is the open-loop gain of the
comparator (practically very high) and Vm is
the biasing voltage of the comparator.
Writing node equation for the node ‘A’ in
fig. 1, we get
𝐶1𝑑𝑢1
𝑑𝑡=
𝑥1
𝑅11+
𝑥12
𝑅21− 𝑢1
1
𝑅11+
1
𝑅21+
1
𝑟1
...(3)
Using equation (1) in (3), we get
𝐶1𝑑𝑢1
𝑑𝑡=
𝑉𝑚 tanh 𝛽(𝑎11𝑉1+ 𝑎12𝑉2 − 𝑏1)
𝑅11+
𝑉𝑚 tanh 𝛽(𝑎21𝑉1+ 𝑎22𝑉2 − 𝑏2)
𝑅21−
𝑢1
𝑅1
...(4)
where 𝑅1 = 𝑅11 𝑅21 𝑟1 ...(5)
And similarly
𝐶2𝑑𝑢2
𝑑𝑡=
𝑉𝑚 tanh 𝛽(𝑎11𝑉1+ 𝑎12𝑉2 − 𝑏1)
𝑅12+
𝑉𝑚 tanh 𝛽(𝑎21𝑉1+ 𝑎22𝑉2 − 𝑏2)
𝑅22−
𝑢2
𝑅2
...(6)
where 𝑅2 = 𝑅12 𝑅22 𝑟2 ...(7)
Moreover, it can be shown that the Energy
Function ‘E’ of the network of fig. 1 is
related to the two voltages V1 and V2 by
𝜕𝐸
𝜕𝑉1 = 𝑉𝑚 𝑎11tanh 𝛽(𝑎11𝑉1 + 𝑎12𝑉2 − 𝑏1) +
𝑉𝑚 𝑎21tanh 𝛽(𝑎21𝑉1 + 𝑎22𝑉2 − 𝑏2) ...(8)
𝜕𝐸
𝜕𝑉2 = 𝑉𝑚 𝑎12tanh 𝛽(𝑎11𝑉1 + 𝑎12𝑉2 − 𝑏1) +
𝑉𝑚 𝑎22tanh 𝛽(𝑎21𝑉1 + 𝑎22𝑉2 − 𝑏2) ...(9)
Also, if ‘E’ is the Energy function, it must
satisfy the following conditions [9].
𝜕𝐸
𝜕𝑉1 = 𝐾 𝐶1
𝑑𝑢1
𝑑𝑡 ...(10)
and
𝜕𝐸
𝜕𝑉2 = 𝐾 𝐶2
𝑑𝑢2
𝑑𝑡 ...(11)
where ‘K’ is a constant of proportionality
and has the dimensions of resistance.
Comparing equations (8) & (10), and (9) &
(11) we get
𝑎11 = 𝐾
𝑅11
𝑎12 = 𝐾
𝑅12
...(12)
𝑎21 = 𝐾
𝑅21
𝑎22 = 𝐾
𝑅22
The values of the resistances of fig. 1 can be
easily calculated by choosing a suitable
value of ‘K’ and then using (12).
Although fig. 1 shows the proposed circuit
as applicable to a system of simultaneous
linear equations in two variables, the
technique can be readily extended to solve
n- simultaneous linear equations in n-
variables.
THE PROCEEDINGS OF IEEE SPONSORED INTERNATIONAL CONFERENCE RASIET-07
IET ALWAR (RAJ) ON 22 -23 DEC 07/VOL.2
R1
R2
R3
R4
V1
V2
R8
R9R
5
R6
R7
2V1+V
2
5V
1V
R10
R11
R12
V1
- V2
C
D
Fig. 2 The proposed network to solve a sample system of linear equations in two variables.
III. CIRCUIT IMPLEMENTATION AND
SIMULATION RESULTS
This section deals with the application of the
network proposed in the previous section to
a sample problem. Let the simultaneous
linear equations to be solved be
2𝑉1 + 𝑉2 − 5 = 0
𝑉1 − 𝑉2 − 1 = 0
For the chosen set of equations in two
variables (V1 and V2), the circuit of fig. 1
transforms to the circuit presented in fig. 2
above. As shown, some additional circuitry
is needed to generate the inputs to the non-
linear synapses.
The values of the resistances in the circuit of
fig. 2 can be obtained by making note of the
fact that the required signal voltages at the
points marked ‘C’ and ‘D’ are (2V1+V2) and
(V1-V2) respectively. Solving for R5 through
R12, we get
R6 = R7 = R8 = R9 = R10 = R11 = R12 ...(14)
and
2R5 = R7 ...(15)
Choosing R6 = R7 = R8 = R9 = R10 = R11 =
R12 = 3K, we get R5 = 1.5 K.
The values of resistances acting as the
weights on the neurons are obtained from
(12). For the purpose of simulation, the
value of ‘K’ was chosen to be 3K. This was
done to have as fewer different valued
resistances as possible. Using K = 3K in
(12) gives
R12 = R21 = R22 = 3K , and
R11 = 1.5K
The results of SPICE simulation for the
circuit of fig. 2 are presented in fig. 3. From
the plots of the voltages at nodes 1 & 2, it
can be seen that V(1) = 2V and V(2) = 1V.
...(13)
THE PROCEEDINGS OF IEEE SPONSORED INTERNATIONAL CONFERENCE RASIET-07
IET ALWAR (RAJ) ON 22 -23 DEC 07/VOL.2
Fig. 3 Simulation results for the circuit of fig. 2
Therefore, the solutions to the simultaneous
linear equations of (13) are V1 = 2V and V2
= 1V. Routine mathematical analysis of (13)
is found to yield the same results.
IV. CONCLUSIONS
A novel neural network based architecture is
presented and utilized to solve simultaneous
linear equations of n- variables. The circuit
thus obtained was realized and simulated
using SPICE for an example set of two
linear equations in two variables. The results
obtained as a result of simulation are found
to verify the theory and are in complete
agreement with the exact mathematical
solution.
REFERENCES
[1] Leonardo V. Ferreira, Eugenius
Kaszkurewicz, and Amit Bhaya, Solving Systems
of Linear Equations Via Gradient Systems With
Discontinuous Righthand Sides: Application to
LS-SVM, IEEE Transactions on Neural
Networks, vol. 16, no. 2, Mar 2005.
[2] Vincent C. Wilburn, Hak-Lim Ko and
Winser E. Alexander, An Algorithm and
Architecture for the Parallel Solution of Systems
of Linear Equations, Fifteenth IEEE Annual
International Phoenix Conference on Computers
and Communications, 1996.
[3] Kechen Zhang, Giorgio Ganis and Martin I.
Sereno, Anti-Hebbian synapses as a linear
equation solver, 1997.
[4] Y. Zhang, D. Jiang and J. Wang, A recurrent
neural network far solving sylvester equation
with time-varying coefficients, IEEE
Transactions on Neural Networks, vol. 13, no.
5, 2002.
[5] C. A. Mead, Analog VLSI and Neural
Systems, Addison-Wesley, United States of
America, 1989.
[6] A. Cichocki and R. Unbehauen, Neural
network for solving systems of linear equations
and related problems, IEEE Transactions on
Circuits and Systems, vol. 39, 1992.
[7] H.T. Kung, Why Systolic Architectures? ,
Computer, vol. 15, January 1982.
[8] S.T. Alexandre, W.E. Alexander, and D.S.
Reeves, An architecture and a simulatior for
parallel signal processing, IEEE Southeastcon
'93 Proceedings, 1993 [9] S. A. Rahman, A Non-linear Synapse Neural
Network and its Applications, Ph. D. Thesis,
2007
Mohd. Samar Ansari received the B.Tech. degree in
Electronics engineering from the Aligarh Muslim
University, Aligarh, India, in 2001 and the M. Tech.
degree in Electronics engineering, with specialization
in Electronic Circuit and System Design, in 2007. He is
currently a Lecturer in the Department of Electronics
Engineering, Aligarh Muslim University, where he
teaches Electronic Devices & Circuits and
Microelectronics. His research interests include microelectronics, linear
integrated circuit applications, neural networks, and semiconductor physics.
Syed Atiqur Rahman did B.Sc.(Engg.) in Electrical Engineering in 1988,
and M.Sc.(Engg.) in Electronics and Communication Engineering in 1994,
from Aligarh Muslim University, Aligarh. He obtained his PhD from
Electrical Engineering Department at IIT (Delhi) in 2007. He is currently a
Reader in the Department of Electronics Engineering, Aligarh Muslim
University, where he teaches neural networks and signals & systems. His
fields of interest are Electronic Circuits and Artificial Neural Networks.
Time
0s 0.2ms 0.4ms 0.6ms 0.8ms 1.0ms 1.2ms 1.4ms 1.6ms 1.8ms 2.0ms
V(1) V(2)
0V
1.0V
2.0V
2.5V
978-1-4244-8542-0/10/$26.00 ©2010 IEEE
Employing Differential Voltage Current Conveyor
in Graph Coloring Applications
Mohd. Samar Ansari
Abstract—A non-linear feedback neural network for solving
graph coloring problem is presented. The proposed circuit
employs non-linear feedback, in the form of unipolar
comparators realized using DVCCs and diodes, to introduce
transcendental terms in the energy function ensuring fast
convergence to the solution. PSPICE simulation results on
various random graphs have been presented.
Index Terms— Differential Voltage Current Conveyor
(DVCC), Dynamical Systems, Feedback Neural Networks, Graph
Coloring, Neural Networks, Non – Linear circuits.
I. INTRODUCTION
he graph coloring problem (GCP) has a large number of
applications such as register allocation in digital
computers, frequency channel assignment in mobile
978-1-4244-8542-0/10/$26.00 ©2010 IEEE
x1Z
+
XY2
Y1
DVCC-1
x1
Ix2
Z+
Y2
Y1
DVCC-2
+
_Vi
Rpi Ci
1/gDNi
X
x1
Ixn
Z+
Y2
Y1
DVCC-n
X
V1
V2
Vn
Ix1
Ixn
Ix2
Ix1
Fig. 2 i-th neuron of the proposed non-linear feedback neural circuit to color a
n-node graph.
As can be seen from fig. 2, individual equations from the set
of constraints are passed through non-linear synapses which
are realized using DVCC-based unipolar comparators. The
outputs of the comparators are fed to neurons having weighted
inputs. These weighted neurons are realized by using opamps
where the currents coming from various comparators act as
weights. Weighing is also provided by the self-feedback
resistance (1/gmD). Node equation for node „Ni‟ gives the
equation of motion of the i-th neurons as
∑
[
] (5)
Where, represent the currents after passing through
diodes as shown in Fig. 2.
Substituting the value of results in
∑ *
( – ) +
[
]
(6)
Moreover, it can be shown that the network in fig. 2 can be
associated with an Energy Function „E‟ of the form
∑
∑ ∑ ( ( ))
∑ ∑
∑
∫
(7)
The last term in (7) is usually neglected for high values of the
open-loop gain of the opamp used to realize the neurons. The
first term on the right hand side of (7) is quadratic which tries
to minimize the number of colors. The second term has got a
negative sign. Therefore, the energy function (E) will be
minimized if second term is maximized. This happens when
the voltages corresponding to connected nodes in a graph are
far away from each other. The first two terms on the right
hand side are balancing each other to color a graph properly.
The third term also contributes to lowering of number of
different colors by eliminating all those local minima in the
energy function for which node voltages are negative.
III. SIMULATION RESULTS
The proposed network was tested for various random graphs
using PSPICE simulations. For the opamp, use was made of
the LMC7101A CMOS opamp from National Semiconductor.
The sub-circuit file for this opamp is available in Orcad Model
Library. Similarly, for the diode the model for D1N3063 diode
available in Orcad Library was utilized. The circuit for DVCC
was taken from [7] and is reproduced in figure 3. Standard 0.5
micron CMOS parameters were used for simulation purposes.
The value of for the DVCC was measured to be 1.091 milli-
mhos.
Y1Y2
VBB
M1 M4M2 M3
M5 M6
M9 M10
M7
M11
M8
M12
XZ+
VSS
VDD
M16
Z-
M17 M18
M15M14M13
Fig. 3 CMOS implementation of DVCC [7]
The results of the tests are given in table-I from which it is
seen that the proposed network gives a solution to all the
problems tested and in all the cases the solution is very near to
the chromatic number of the graph. The results of PSPICE
simulations for graph-3 of table-I are presented in figure 4
from which it can be seen that the proposed circuit assigns
only three different voltages to the different nodes.
Table I PSPICE simulation results of the proposed circuit applied
to color different graphs
S. No.
Test Graph
Simulation Results
Chromatic Number No. of
colors
Frequency of
occurrence
1
2 7/10 2
3 3/10
2
2 4/10
2
3 6/10
3
3 10/10 2
978-1-4244-8542-0/10/$26.00 ©2010 IEEE
IV. CONCLUSIONS
In this paper, a CMOS compatible approach to solve graph
coloring problems is presented. The proposed approach
employs n-neurons and n(n-1)-synapses for a n-node graph to
be colored. Each neuron requires one opamp and each synapse
is comprised of DVCCs. This results in significant reduction
in hardware over the existing schemes [x-y]. Specifically the
number of passive floating resistances is limited to one per
neuron. The proposed network was tested on 3 random graphs
and the simulation results confirm the validity of the approach.
V. REFERENCES
[1] T. Zeitlhofer and B. Wess, “A Comparison of Graph Coloring Heuristics
for Register Allocation based on Coalescing in Interval Graphs,” Proc. ISCAS 04, vol. IV, pp. 529 – 532, 2004.
[2] N. A. El-Fishawy, M. M. Hadhood, S. Elnoubi, and W. EL-Sersy, “A
modified Hopfield neural network algorithm for cellular radio channel assignment,” Proc. TENCON 2000, vol. 2, pp. 213 – 216, Sept 2000.
[3] A. D. Blas, A. Jagota and R. Hughey, “Energy Function – Based
Approaches to Graph Coloring,” IEEE Trans. On Neural Networks, vol.
13, no. 1, pp. 81 – 91, Jan 2002.
[4] J. Lui, W. Zhong, and L. Jiao, „Comments on “The 1993 DIMACS
Graph Coloring Challenge” and “Energy Function-Based Approaches to Graph Coloring”,‟ IEEE Trans. On Neural Networks, vol. 17, no. 2, pp.
533, Mar 2006. [5] S. A. Rahman, Jayadeva and S. C. Dutta Roy, “Neural network approach
to graph colouring,” Electronics Letters, vol. 35, no. 14, pp. 1173 –
1175, Jul 1999 [6] C. W. Wu, “Graph Coloring via Synchronization of Coupled
Oscillators,” IEEE Trans. Circuits Syst. I, vol. 45, no. 9, pp. 974 – 978,
Sept 1998. [7] S. Maheshwari, “A canonical Voltage-Controlled VM-APS with a
Grounded Capacitor,” Cir. Sys. Sig. Processing, vol. 27, pages 123-132,
2008.
Fig. 4 PSPICE simulation results for graph-3 in Table-I
Time
0s 10us 20us 30us 40us 50us 60us 70us 80us 90us 100us
V(1) V(2) V(3) V(4) V(5) V(6) V(7) V(8) V(9) V(10)
-5V
0V
5V
10V
15V
International Conference on Intelligent Systems and Networks (ISN-2008), February 22-24, 2008.
Computational Intelligence Laboratory, Institute of Science and Technology, Klawad-133105, Haryana, India
123
Solution of Quadratic Programming Problem
using Feedback Neural Network
M. S. Ansari1 and S. A. Rahman
2
1,2 Department of Electronics Engineering, Z. H. College of Engineering & Technology,
Aligarh Muslim University, Aligarh-202002, India. 1Tel: +91-9927014450, E-mail: [email protected]
2Tel: +91-571-3204389, E-mail: [email protected]
Abstract—A large number of optimization problems, having
non-linear objective functions, are usually approximated by
a second order form. Solving a quadratic programming
problem (QPP) efficiently is one of the most important tasks
in optimization techniques. This is because the solutions of a
QPP using traditional methods include the use of parallel
computing architectures and complex software algorithms.
The time taken by the existing methods to solve a
constrained optimization problem is not always short and
increases significantly with the number of variables. This
paper presents a feedback neural network based hardware
solution to the problem of minimizing a second-order
objective function subject to linear constraints. A direct
consequence of the use of neural networks is massively
parallel processing and fast convergence. This results in a
faster and more compact hardware. Simulation results are
presented for a chosen problem and are found to agree
exactly with the algebraic solution.
I. INTRODUCTION
In the discipline of constrained optimization,
linear programming has been studied the most because of
its fundamental nature and wide applicability. However,
another class of constrained optimization problem is the
quadratic programming problem (QPP). It deals with the
optimization (minimization or maximization) of a second-
order objective function subject to certain linear
constraints. Optimization problems with higher order,
non-linear objective functions are usually approximated
by a second-order form and solved approximately by
standard quadratic programming techniques. Traditional
techniques to solve a QPP normally involve an iterative
process. Due to this iterative nature of the solution
process, several computer algorithms were proposed.
However, the problems to be solved often are of very
large size, so that large computer resources or massive
parallel computer systems with distributed memory are
needed to solve them. When the problem at hand is large,
we must obtain its solution within a reasonably short
period of time. Existing methods, for solving a quadratic
programming problem, include Newton-Barrier method
and sequential quadratic programming (SQP). For all
these methods, the time taken to solve a QPP increases
significantly with the number of variables.
This paper presents a hardware solution to the
problem of solving a quadratic programming problem.
The approach used is to employ neural networks on the
basis of an analog circuit [1-6]. The most important
advantages of neural networks are massively parallel
processing and fast convergence. Hence, architectures
based on neural networks have many computational
advantages over the traditional algorithms running over a
parallel computing system.
The following section lays the basic problem and
the mathematical formulation on which the development
of the proposed network will be based. Section III
contains the circuit implementation of the proposed
network for a set of sample problem in two variables.
SPICE simulation results of the proposed circuit are also
presented. Some concluding remarks are presented in
Section IV.
II. PROPOSED NEURAL NETWORK BASED CIRCUIT
Let the objective function to be minimized be
... (1)
subject to the constraints
where V1 and V2 are the two variables and c11, c12, c22, a11,
a12, a21, a22, b1 and b2 are constants.
The proposed neural-network based circuit to
solve the quadratic programming problem given above is
presented in fig. 1. As can be seen from fig. 1, individual
inequalities from the set of constraints are passed through
a non-linear synapse. Each non-linear synapse is in effect
a unipolar comparator, which is realized using an opamp
and a diode as shown. The outputs of the non-linear
synapses are fed to weighted neurons. These weighted
neurons are again realized by using Opamps where the
resistors R11 through R22, and Rc11 through Rc22, act as
weights. Rpi and Cpi are the parasitic resistance and
capacitance that are present at the input of the opamp used
in the realization of the i-th neuron. These parasitic
components are included to model the dynamic nature of
the Opamp.
Non-linear synapses are realized using
comparators. The comparator outputs can be modelled by
equations (4) and (5) given below.
(2)
(3)
International Conference on Intelligent Systems and Networks (ISN-2008), February 22-24, 2008.
Computational Intelligence Laboratory, Institute of Science and Technology, Klawad-133105, Haryana, India
124
..(4)
..(5)
where β is the open-loop gain of the comparator
(practically very high) and Vm is the biasing voltage of the
comparator.
Writing node equation for the node „A‟ in fig. 1, we get
...(6)
Where
..(7)
Similarly, for node „B‟ in figure – 1, we get
...(8)
where
..(9)
Moreover, it can be shown that the Energy
Function „E‟ of the network of fig. 1 is related to the two
voltages V1 and V2 by
...(10)
(11)
Also, if „E‟ is the Energy function, it must satisfy the
following conditions [1, 2, 6]
...(12) And
...(13)
where „K‟ is a constant of proportionality and
has the dimensions of resistance.
Comparing equations (10) & (12), we get
...(14)
A similar comparison between (11) & (13) yields
...(15)
The values of the resistances of fig. 1 can be
easily calculated by choosing a suitable value of „K‟ and
then using (14, 15).
Although fig. 1 shows the proposed circuit as
applicable to a second-order constrained minimization
problem in two variables, the technique can be readily
extended to solve similar problems for higher number of
variables.
III. CIRCUIT IMPLEMENTATION AND SIMULATION
RESULTS
This section deals with the application of the network
proposed in the previous section to a sample problem. Let
the objective function to be minimized be
...(16)
International Conference on Intelligent Systems and Networks (ISN-2008), February 22-24, 2008.
Computational Intelligence Laboratory, Institute of Science and Technology, Klawad-133105, Haryana, India
125
subject to
...(17)
The values of resistances acting as the weights
on the neurons are obtained from (14, 15). For the purpose
of simulation, the value of „K‟ was chosen to be 10 KΩ.
Using K = 10 K in (12) gives
Rc11 = Rc11 = Rc11 = Rc11 = K = 10 KΩ
R11 = 1.66 KΩ
R22 = 1 KΩ
R12 = R21 = 2.5 KΩ
Fig. 2 Simulation results for the circuit of fig. 1 applied to minimize
(16,17)
The results of SPICE simulation are presented in fig. 2.
From the plots of the neuron output voltages, it can be
seen that V(1) = -0.58 V and V(2) = 0.41 V. Routine
mathematical analysis of (16) is found to yield the same
results.
IV. CONCLUSIONS
A new neural network based architecture is
presented and utilized to solve a quadratic optimization
problem subject to linear constraints. The circuit thus
obtained was realized and simulated using SPICE for a
sample minimization problem in two variables. The
results obtained from simulation are found to verify the
theory and are in complete agreement with the exact
mathematical solution.
REFERENCES
[1] S. A. Rahman and Mohd. S. Ansari, “Solution of imultaneous Linear Equations using Feedback Neural Network,” Proc.
RASIET-07, International Conference on Recent Applications of
Soft Computing in Engineering and Technology, IET, Alwar, India.
[2] S. J. Arif, S. A. Rahman and Mohd. S. Ansari, “A Non – Linear
Feedback Neural Network for Graph Coloring,” Proc. RASIET-07, International Conference on Recent Applications of Soft
Computing in Engineering and Technology, IET, Alwar, India.
[3] F.L. Filippelli, M. Forti and S. Manetti, “New linear and quadratic programming neural network,” Electronics Letters, vol. 30, no. 20,
September 1994.
[4] M. P. Kennedy, and L. O. Chua, "Neural networks for nonlinear programming," IEEE Trans. 1988, CAS-35, pp. 554-562.
[5] C. Y. Maa, and M. A. Shanblatt, "Linear and quadratic
programming neural network analysis," IEEE Trans., 1992. NN-4, pp. 580-594.
[6] S. A. Rahman, Jayadeva and S. C. Dutta Roy, “Neural network approach to graph colouring,” Electronics Letters, vol. 35, Jul
1999
Fig. 1 The proposed Neural Network based circuit to solve a quadratic programming problem in 2 variables
IMPACT-2009
A non-linear neural circuit for solving system of
simultaneous linear equations
Mohd. Samar Ansari1 and Syed Atiqur Rahman
2
1,2Department of Electronics Engineering, Aligarh Muslim University, Aligarh, India
1Email:[email protected],
2Email:[email protected]
Abstract— A feedback neural network based CMOS
compatible circuit to solve a system of simultaneous linear
equations is presented. The circuit has an associated
transcendental energy function that ensures fast convergence to
the exact solution while enjoying reduction in hardware
complexity over existing schemes. HSPICE simulation results
are presented for a chosen set of equations and are found to
agree exactly with the algebraic solution.
Index Terms—Neural network applications, Neural network
hardware, Nonlinear circuits, linear algebra, linear equations
I. INTRODUCTION
HE solution of linear equations has been a primary goal
for computation since the time of the abacus. References
to word problems requiring the solution of a system of linear
equations appear in ancient Babylonian texts dating back to
circa 300 BC [1]. Some examples of applications requiring
the solution of linear equations are curve fitting, electrical
circuit analysis, multiple correlation as well as real time
applications like real-time speech coding, image processing,
stochastic modeling, and computer-aided realistic three-
dimensional image synthesis [1-9]. Babbage‟s Analytical
Engine (1836) was the first attempt at automating the
equation solving process. It was followed by the Atanasoff-
Berry Computer (1941) which marked the transition from
mechanical to an electronic computing architecture [1, 13].
However, consuming almost a minute for each
addition/multiplication, the Atanasoff-Berry Computer was
certainly slow. The era of fast equation solving computers
was signaled by FPS-164/MAX (1984). It was followed by
the ClearSpeed CSX600 (2005) which was capable of
working at 25 Gflops/s consuming 10 Watts [1]. However,
such architectures as the ClearSpeed CSX600 or the FPS-
164/MAX employ massively parallel digital hardware
employing several coprocessors and running a full blown
operating system. Such an arrangement may not be suitable
for real-time and/or portable applications where a dedicated,
compact, low-power solution is desirable. This has led to
research efforts being directed towards the development of
specialized hardware for the solution of linear equations. The
dynamical system approach was first used by Pyne to solve a
system of linear equations [2]. Later, neural networks
promising massively parallel processing and fast convergence
were also applied to solve linear equations [3, 4, 5].
In this paper, a new non-linear feedback neural architecture
with reduced circuit complexity an energy function involving
transcendental terms which is fundamentally different from
the standard quadratic form associated with Hopfield network
and its variations. Further, from the point of view of
implementation in VLSI technology, the proposed circuit
uses a multi-output transconductance cell in conjunction with
an operational amplifier to emulate a weighted neuron unlike
its predecessor where resistances were utilized to realize the
weights to the neurons [9].
This paper is organized as follows. A brief review of
existing methods for solving linear equations appears in
section – II. Section – III contains the details of the proposed
network along with the design equations. Proofs of the energy
function and validity of the solution are given in section – IV.
Section – V presents the results of HSPICE simulation of the
circuit applied to solve a sample equation set. A discussion
on VLSI implementability of the proposed circuit appears in
section – V. Some conclusive remarks appear in section – VI.
II. EXISTING METHODS
Existing methods to solve a set of linear equations
manually include Gaussian elimination and the Cholesky
decomposition [9, 12]. To determine approximate solutions to
general systems numerically using a software program
running on a computer, the n-dimensional Newton's method
may be used [12]. Software linear equation solving programs
like „SLAE Solver‟ and „Sogol‟ are also commercially
available [12].However, all these methods, being sequential
in nature, are not suitable for real-time processing.
Hardware solutions based on neural networks have been
put forward by Xia, Wang & Hung [5], Cichocki &
Unbehauen [6] and Wang [7]. To solve an n-variable system
of equations, the network of [7] uses three operational
amplifiers, one capacitor and (n+5) resistances to emulate a
single neuron and the time to arrive at the solution is of the
order of hundreds of milliseconds. The network of [6]
provides a significantly improved solution time of around a
microsecond but at the cost of increased hardware
complexity. Each neuron in [6] comprises of three weighted
summers and an inverting integrator. The architecture of [5],
which is a generalized neural network based implementation
of Censor and Elfving’s method for linear inequalities, also
uses an approach similar to [7] utilizing weighted adders and
integrators to realize the neurons.
T
978-1-4244-3604-0/09/$25.00 ©2009 IEEE 120
IMPACT-2009
Ni
Among the current state of the art solutions, the recently
developed concurrent multiprocessor based architectures for
solving linear equations; only systolic/wavefront arrays and
the Block Data Parallel Architecture (BDPA) are suited for
solving computationally intensive problems [9]. However,
even with the extensive simulation tools developed for the
BDPA, there is still a fundamental need to show the
prospective power of such architectures. The ClearSpeed
CSX600 (2005) capable of solution times of the order of
microseconds is among the fastest available linear equation
solver. However, it needs to run a host operating system and
is prone to “OS jitter” at high operating speeds [1]. These
issues together with the fact that its power consumption is
around 10 Watts make the CSX600 unsuitable for real-time
and/or portable applications.
The aim of this paper is to provide a CMOS compatible
neural circuit, the circuit complexity of which compares
favorably with existing networks while enjoying very fast
convergence time.
III. PROPOSED NEURAL NETWORK
Let a system of linear equations in n- variables be
a11 a12 … a1n
a21 a22 … a2n
⋮ ⋮ … ⋮an1 an2 … ann
𝑉1
𝑉2
⋮𝑉𝑛
=
b1
b2
⋮bn
(1)
Where V1, V2,..., Vn are the variables and aij and bi are
constants. The proposed neural-network based circuit to solve
the system of equations of (1) is presented in fig. 1.
As can be seen from fig. 1, individual equations from the
set of equations to be solved are passed through non-linear
synapses which are realized using comparators. The output of
the i-th voltage-mode comparator can be modelled by
𝑥𝑖 = 𝑉𝑚 𝑡𝑎𝑛ℎ 𝛽 𝑎𝑖1𝑉1 + 𝑎𝑖2𝑉2 + ⋯ + 𝑎𝑖𝑛𝑉𝑛 − 𝑏𝑖 (2)
Where β is the open-loop gain of the comparator (practically
very high), ±Vm are the output voltage levels of the
comparator and V1, V2,..., Vn are the neuron outputs.
Vi
Vout+
_
Vi
Vout+
_+
_V
i
b1
b2
x1
x2
Ri
Ci
Vi
Vout+
_
bn x
n
G1
G2
Gn
a11
V1+ a
12V
2 + ...+a
1nV
n
a21
V1+ a
22V
2 +...+ a
2nV
n
an1
V1+ a
n2V
2 + ...+a
nnV
n
g1i
x1
g2i
x2
gni
xn
To other
neurons
Fig. 1 i-th neuron of the proposed feedback neural network circuit to solve
simultaneous linear equations in n-variables
The outputs of the comparators are fed to neurons having
weighted inputs. The neurons are realized by using opamps
and the weights are implemented using transconductance
amplifiers. The currents arriving at the input of a neuron
from various synapses are added using KCL. Rpi and Cpi are
the input resistance and capacitance of the opamp
corresponding to the i-th neuron. These parasitic components
are included to model the dynamic nature of the opamp.
Node equation for node „Ni‟ gives the equation of motion of
the i-th neuron as
𝐶𝑖𝑑𝑢𝑖
𝑑𝑡= 𝑔1𝑖𝑥1 + 𝑔2𝑖𝑥2 + ⋯ + 𝑔𝑛𝑖𝑥𝑛 −
𝑢𝑖
𝑅𝑖 (3)
Where ui is the internal state of the i-th neuron and gji is the
voltage-to-current conversion factor of the j-th transconductance block for the i-th output current. As is shown later in this section, these weights are governed by the entries in the coefficient matrix of (1).
Using (2) in (3) results in
𝐶𝑖𝑑𝑢𝑖
𝑑𝑡= 𝑔1𝑖𝑉𝑚 tanh 𝛽(𝑎11𝑉1 + 𝑎12𝑉2 + ⋯ + 𝑎1𝑛𝑉𝑛 − 𝑏1) +
𝑔2𝑖𝑉𝑚 tanh 𝛽(𝑎21𝑉1 + 𝑎22𝑉2 + ⋯ + 𝑎2𝑛𝑉𝑛 − 𝑏2) + ⋯ +
𝑔𝑛𝑖𝑉𝑚 tanh 𝛽(𝑎𝑛1𝑉1 + 𝑎𝑛2𝑉2 + … + 𝑎𝑛𝑛 𝑉𝑛 − 𝑏𝑛) − 𝑢𝑖
𝑅𝑖 (4)
Moreover, as has been shown in section-IV, the network in
fig. 1 can be associated with an Energy Function „E‟ of the
form
𝐸 =
𝑉𝑚 ln cosh 𝛽 𝑎𝑖𝑗𝑛𝑗=1 𝑉𝑗 − 𝑏𝑖 –
1
𝑅𝑖
𝑛𝑖=1 𝑛
𝑖=1 𝑢𝑖𝑉𝑖
0𝑑𝑉𝑖 (5)
From (5), it follows that
𝜕𝐸
𝜕𝑉1= 𝑉𝑚 𝑎11tanh 𝛽(𝑎11𝑉1 + 𝑎12𝑉2 + ⋯ + 𝑎1𝑛𝑉𝑛 − 𝑏1) +
𝑉𝑚𝑎21tanh 𝛽(𝑎21𝑉1 + 𝑎22𝑉2 + ⋯ + 𝑎2𝑛𝑉𝑛 − 𝑏2) + ⋯ +
𝑉𝑚 𝑎𝑛1tanh 𝛽(𝑎𝑛1𝑉1 + 𝑎𝑛2𝑉2 + ⋯ + 𝑎𝑛𝑛𝑉𝑛 − 𝑏𝑛) − 𝑢1 1
𝑅1 (6)
Also, if „E‟ is the Energy Function, it must satisfy the
following condition [10].
𝜕𝐸
𝜕𝑉1= 𝐾 𝐶1
𝑑𝑢1
𝑑𝑡 (7)
Where „K‟ is a constant of proportionality and has the
dimensions of resistance. Comparing (4) and (6) according to
(7) yields
𝑎11 = 𝐾𝑔11 , 𝑎21 = 𝐾𝑔11 , … , 𝑎𝑛1 = 𝐾𝑔11 (8)
The values of the transconductance factors (gji) of fig. 1
can be easily calculated by choosing a suitable value of „K‟
and then using (8). Analysis on similar lines can be
performed to obtain the values of the synaptic weights for the
remaining neurons.
121
IMPACT-2009
IV. ENERGY FUNCTION
Differentiating the energy function of (5) w.r.t. Vi, we get
𝜕𝐸
𝜕𝑉𝑖= 𝑉𝑚𝑎1𝑖 tanh 𝛽(𝑎11𝑉1 + 𝑎12𝑉2 + ⋯ + 𝑎1𝑛𝑉𝑛 − 𝑏1) +
𝑉𝑚𝑎2𝑖 tanh 𝛽(𝑎21𝑉1 + 𝑎22𝑉2 + ⋯ + 𝑎2𝑛𝑉𝑛 − 𝑏2) + … +
𝑉𝑚𝑎𝑛𝑖 tanh 𝛽(𝑎𝑛1𝑉1 + 𝑎𝑛2𝑉2 + … + 𝑎𝑛𝑛 𝑉𝑛 − 𝑏𝑛) − 𝑢𝑖
𝑅𝑖 (9)
Comparing (4) and (6), we observed that
𝜕𝐸
𝜕𝑉𝑖= 𝐾𝐶𝑖
𝑑𝑢𝑖
𝑑𝑡 (10)
Where
𝐾 = 𝑎1𝑖/𝑔1𝑖 ; for all i
Further, the time derivative of the energy function is given by
𝑑𝐸
𝑑𝑡=
𝜕𝐸
𝜕𝑉𝑖
𝑑𝑉𝑖
𝑑𝑡
𝑁𝑖=1 =
𝜕𝐸
𝜕𝑉𝑖
𝑑𝑉𝑖
𝑑𝑢𝑖
𝑁𝑖=1
𝑑𝑢𝑖
𝑑𝑡 (11)
Using (10) in (11) we get
𝑑𝐸
𝑑𝑡= 𝐾𝐶𝑖
𝑁𝑖=1
𝑑𝑢𝑖
𝑑𝑡
2 𝑑𝑉𝑖
𝑑𝑢𝑖 (12)
For the inverting opamp used in the realization of the neuron,
the transfer characteristics are shown in fig. 2. From these
characteristics, it is evident that
𝑑𝑉𝑖
𝑑𝑢𝑖 ≤ 0 (13)
Thereby resulting in
𝑑𝐸
𝑑𝑡≤ 0 (14)
With the equality being valid for
𝑑𝑢𝑖
𝑑𝑡= 0 ; for all i (15)
Equation (14) shows that the energy function can never
increase with time which is one of the conditions for a valid
energy function. The second criterion viz. the energy function
must have a lower bound is also satisfied for the circuit of
fig.1 wherein it may be seen that V1, V2,..., Vn are all
bounded (being the outputs of opamps) amounting to „E‟
having a definite lower bound.
Fig. 2 Transfer characteristics of the opamp used to realize the neurons
The plot of the energy function for a 2 – variable problem
is shown in fig. 3. From the plot, it is evident that there exists
only one minimum to which the network must converge,
which is the solution of the given system of linear equations.
Fig. 3 Typical energy function plot for a 2 – variable problem
V. SIMULATION RESULTS
The proposed circuit was simulated for solving the
following set of 2 simultaneous linear equations in 2
variables.
2 11 1
V1
V2 =
21.5
(16)
The circuit to solve (16) as obtained from fig. 1 is presented
in fig. 4. As can be seen, some additional circuitry is needed
to generate the inputs to the non-linear synapses. Routine
analysis yields the following values of the resistors at the
input of the non-linear synapses.
R11=2K, R12=4K, R13=4K, R21=3K, R22=3K and R23=3K.
Vi
Vout+
_
Vi
Vout+
_+
+
_
_
V1
V2
b1
b2
x1
x2
Rp1
Rp2
C1
C2
G1
G2
R11
R12 R
13
R21
R22 R
23
a11
x1
a21
x1
a12
x2
a22
x2
Fig. 4 The proposed circuit applied to a 2 – variable problem
For the purpose of HSPICE simulations, the voltage
comparator and the multi-output transconductance amplifier
for each neuron were realized as a single block using a
slightly modified circuit of the differential-input, high-gain,
active-loaded transconductance amplifier [11] shown in fig.
5. Standard BSIM3 0.35µm parameters were used for the
purpose of simulations. The supply voltages were set to VDD
= 15V, VSS = -15V and VBB = -14V. The aspect ratios of the
NMOS and PMOS transistors were taken to be 1.4µm/1.4µm
and 2.8µm/1.4µm respectively. Further, to get the output
currents according to (8) the W/L ratios of transistors M6,
M7, M8 and M9 were set to provide the required current
scaling. Choosing K=1 and using (8) results in Iout1 = 2Iout2 for
G1 block and Iout1 = Iout2 for the G2 block of fig. 4.
Routine mathematical analysis of (16) gives the solution as
V1 = 0.5 and V2 = 1. The results of HSPICE simulation of the
circuit of fig. 4, shown in fig. 6, are found to match exactly
with the algebraic solution.
122
IMPACT-2009
M2
M1
M3
M4
M5
M6
M7
VDD
VSS
VBB
V2
V1 I
out1
M8
M9
Iout2
Fig. 5 CMOS implementation of the comparator and transconductance blocks
Fig. 6 Simulation results for the chosen 2 – variable problem
A qualitative analysis of the simulation results of fig. 6
shows that the voltage-mode outputs obtained are indeed the
correct solutions of (16). The HSPICE output file for the
simulation run gives the value of V1 as 0.499V and V2 as
1.002V which are so near to the algebraic solution that for all
practical purposes they may be considered as exact. Another
noteworthy point in fig. 6 is that the time taken by the circuit
to provide the solutions is significantly reduced as compared
to the time taken by existing schemes [5, 6, 7].
VI. DISCUSSION
This section deals with the monolithic implementation
issues of the proposed circuit. The HSPICE simulations
assumed that all operational amplifiers (and transconductance
amplifiers) are identical, and therefore, analysis is required to
determine how deviations from this assumption affect the
performance of the network. Effects of variations in
component values from one neuron to another also need to be
investigated.
Moreover, for obtaining weighted (scaled) current outputs
use was made of the (W/L) ratios of the transistors M6, M7,
M8 and M9 of fig. 5. This results in a particular circuit
implementation „hard-wired‟ to solve a given set of
equations. Methods to scale the currents by a mechanism
external to the circuit need to be investigated to impart
flexibility to the circuit. Alternative realizations based on the
differential equations (4) governing the system of neurons are
also being considered. Other approaches to obtain the tanh(.)
nonlinearity include the use of a MOSFET operated in the
subthreshold region [11] and the use of Current Differencing
Transconductance Amplifier (CDTA) to provide the same
nonlinearity in the current-mode regime.
VII. CONCLUSIONS
In this paper we have described a CMOS compatible
approach to solve simultaneous linear equations in n
variables, which uses n-neurons and n-synapses. Each neuron
requires one opamp and each synapse is implemented using a
differential transconductance element. This results in
significant reduction in hardware over the existing schemes
[5-7]. The proposed network was tested on a sample problem
set of two simultaneous linear equations and the simulation
results confirm the validity of the approach.
REFERENCES
[1] J. L. Gustafson, “The Quest for Linear Equation Solvers and the
Invention of Electronic Digital Computing,” Proc. JVA‟06, IEEE John
Vincent Atanasoff 2006 International Symposium on Modern Computing, 2006
[2] I. B. Pyne, “Linear programming on an electronic analogue computer,”
Trans. Amer. Inst. Elec. Eng., vol. 75, pp. 139-143, May 1956. [3] K. Zhang, G. Ganis and M. I. Sereno, “Anti-Hebbian synapses as a
linear equation solver,” Proc. International Conference on Neural
Networks, 9 – 12 June 1997, vol. 1, Page(s): 387 – 389
[4] D. Jiang, “Analog Computing for Real-Time solution of Time-Varying Linear Equations,” Proc. ICCCAS, Volume 2, 27-29 June 2004.
[5] Y. Xia, J. Wang and D. L. Hung, “Recurrent Neural Networks for
Solving Linear Inequalities and Equations,” IEEE Tran. Cir. & Sys - I, vol. 46, no. 4, Apr 1999.
[6] A. Cichocki and R. Unbehauen, “Neural Networks for Solving Systems
of Linear Equations and Related Problems,” IEEE Tran. Cir. & Sys - I, vol. 39, no. 2, Feb 1992.
[7] J. Wang, “Electronic Realization of Recurrent Neural Network for
Solving Simultaneous Linear Equations,” Electronics Letters, vol. 28, no. 5, Feb 1992.
[8] M.P. Kennedy, L.O. Chua, “Neural networks for nonlinear
programming,” IEEE Trans. Circuits Sys, vol. 35, no. 5, 1988.
[9] S. A. Rahman and M. S. Ansari, “Solution of Simultaneous Linear
Equations using Feedback Neural Network,” Proc. RASIET, IET,
Alwar, India, Dec 2007.
[10] S. A. Rahman, Jayadeva and S. C. Duttaroy, “Neural network approach
to graph colouring,” Electronics Letters, vol. 35, no. 14, July 1999
[11] R. W. Newcomb and J. D. Lohn, “Analog VLSI for neural networks,”
in The Handbook of Brain Theory and Neural Networks. Cambridge,
MA: The MIT Press, 1998.
[12] Web Link: http://en.wikipedia.org/wiki/Simultaneous_linear_equations
[13] S. White, “A Brief History of Computing - Complete Timeline,” Web
link: http://trillian.randomstuff.org.uk/~stephen//history/timeline.html
123
IMPACT-2009
A Novel Current-mode Non-Linear Feedback Neural Circuit for Solving Linear Equations
Mohd. Samar Ansari1 and Syed Atiqur Rahman2
1,2Department of Electronics Engineering, Aligarh Muslim University, Aligarh, India Emails: [email protected], [email protected]
Abstract— A novel current-mode neural circuit employing non-linear feedback to solve a system of simultaneous linear equations is presented. The circuit has an associated transcendental energy function that ensures fast convergence to the exact solution while enjoying a resistor-less implementation. The hardware complexity of the proposed scheme compares favourably with existing voltage-mode neural circuits for the same task. PSPICE simulation results are presented for a chosen set of equations and are found to be in agreement with the algebraic solution. Index Terms—Neural network applications, Current mode neural network, Neural network hardware, Nonlinear circuits, linear equations, Linear algebra
I. INTRODUCTION
OLVING a system of linear equations Ax = b is one of the most fundamental problems in algebra. Countless applications including, but not limited to, real-time speech coding, image processing, multiple correlation and electrical circuit analysis, occur in the mathematical sciences and engineering. [1-5, 10, 11, 12]. Popular methods to solve large-size problems demand massively parallel computer systems with distributed memory [1, 12]. Apart from such software approaches, some hardware solutions based on artificial neural networks (ANNs) have been proposed over the past decade. Massively parallel processing and fast convergence, which are inherent to neural networks, are utilized to reduce the time taken to arrive at the solution [1-3]. However, some existing neural networks have variable penalty parameters which decrease to zero as time increases to infinity in order to get better accuracy of solution [2]. Therefore, the time taken to arrive at the solution in such cases is not always short [2]. Further, to the best of our knowledge, no current-mode (CM) neural network for solving a system of simultaneous linear equations is reported in the technical literature although the advantages of current-mode (CM) processing in ANNs are well documented [8]. One of the most apparent advantages is that the summing of many signals is most readily accomplished when those signals are currents [8]. Other benefits include better noise immunity and the possibility of resistor-less realizations.
A new current-mode architecture with reduced circuit complexity is presented which avoids the use of penalty parameters. The proposed architecture uses non-linear feedback which leads to a new energy function that involves transcendental terms. This transcendental energy function is fundamentally different from the standard quadratic form associated with Hopfield network and its variants. The proposed circuit uses only multi-output Current Differencing Transconductance Amplifiers (CDTAs) [5] to realize weighted neurons unlike its predecessor where opamps were used to realize neurons and resistances were utilized to realize the weights to the neurons [6]. This paper is organized as follows. A brief review of existing VM methods for solving linear equations appears in section – II. Section – III contains the details of the proposed CM network along with the design equations. Section – IV presents the results of SPICE simulation of the circuit. A discussion on VLSI implementability of the proposed circuit appears in section – V. Some conclusive remarks appear in section – VI.
II. EXISTING VOLTAGE-MODE METHODS Voltage-mode hardware solutions based on neural
networks have been put forward by Wang [2], Cichocki & Unbehauen [3] and Xia, Wang & Hung [4]. To solve an n-variable system of equations, the network of [2] uses three operational amplifiers, one capacitor and (n+5) resistances to emulate a single neuron and the time to arrive at the solution is of the order of hundreds of milliseconds. The network of [3] provides a significantly improved solution time of around a microsecond but at the cost of increased hardware complexity. Each neuron in [3] comprises of three weighted summers and an inverting integrator. The architecture of [4], which is a generalized neural network based implementation of Censor and Elfving’s method for linear inequalities, also uses an approach similar to [2] utilizing weighted adders integrators to realize the neurons.
The aim of this paper is to provide a novel current-mode neural circuit for the solution of a set of simultaneous linear equations using only CDTAs. No passive components are needed for the synaptic interconnections and therefore, the
S
978-1-4244-3604-0/09/$25.00 ©2009 IEEE 284
IMPACT-2009 circuit complexity of the proposed scheme compares favorably with existing networks while enjoying very fast convergence time.
III. PROPOSED NEURAL NETWORK
The schematic symbol of the CDTA is shown in figure 1. The inputs (+) and (-) produce difference current which is transferred to the z terminal and the voltage at the z terminal is converted into output current by a multi-output transconductance stage. The port relations characterizing CDTA are given by
(1)
Fig. 1 Schematic symbol of CDTA
For the purpose of use in the proposed neural circuit, only multi-output X+ type CDTAs are needed. Although the schematic in figure 1 shows only one X+ output, the number of such outputs may readily be increased by repeating the output stages in the actual circuit implementation of the CDTA [5]. Let the system of n- simultaneous linear equations in n- variables be characterized by
(2)
Where I1, I2,..., In are the variables and aij and bi are constants. The proposed neural-network based circuit to solve the system of equations of (2) is presented in figure 2. As can be seen from figure 2, individual equations from the set of equations to be solved are passed through non-linear synapses which are realized using current-mode comparators. The unweighted output of the i-th comparator can be modelled by
(3)
Where β is the open-loop gain of the comparator (practically very high), ±Im are the output current levels of the comparator and I1, I2,..., In are the neuron outputs. Equation (3) shows that the synaptic connections are highly non-linear as opposed to the linear synaptic interconnections used in standard Hopfield network and its variants.
The outputs of the comparators are fed to neurons as weighted inputs. Weights are assigned to the output currents obtained from the comparators according to the entries in the coefficient matrix [aij]. The neurons are realized by using CDTAs wherein the currents arriving at the input of a neuron from various synapses are added. It may be mentioned that the circuit of figure 2 is a current-mode version of the i-th neuron of [6] where opamps were used to implement the non-linear synapses and neurons thereby resulting in voltage-mode operation and generating voltage outputs corresponding to the solutions of the set of linear equations.
Fig. 2 i-th neuron of the proposed current-mode non-linear feedback neural network circuit to solve simultaneous linear equations in n-variables The synaptic weights, or equivalently, the output current scaling factors are equal to the entries in the coefficient matrix of (2). Since the outputs of the comparators are available as currents, there is no need of resistors as were used in [6] to convert the comparator output voltage available into currents that were then added at the input of the amplifier acting as the neuron. Another advantage arising due to current-mode operation is the reduction in the overall power consumed by the circuit to arrive at the answers. This is because the CDTAs used in the network of figure 2 work with currents of the orders of microamperes and can operate from a power supply lower than that normally required for opamps. Moreover, the proposed circuit does not use resistances (floating or grounded) thereby yielding a ‘transistor-only’ implementation. This causes the circuit to compare very favourably with existing circuits from the viewpoint of the suitability of the circuit for monolithic VLSI implementation. Further, since the circuit works in a fashion very similar to its voltage-mode predecessor, the associated energy function will continue to be of the same form as was proposed earlier in [6] and is reproduced below for convenience.
(4)
+_
IDiffIout
+_b1
b2
a1ix1
+_bn
a11I1+ a12I2 + ...+a1nIna21I1+ a22I2 +...+ a2nIn
an1I1+ an2I2 + ...+annIn
To other neuronsa2ix2
anixn
IDiff
IDiff
Iout
Iout
+_
0Iia1iIianiIi Feedback CurrentsCDTA
285
IMPACT-2009
Fig. 3 Bipolar realization of CDTA [5]
IV. RESULTS OF COMPUTER SIMULATION The proposed circuit was simulated for solving the following set of 2 simultaneous linear equations in 2 variables.
(5)
The circuit to solve (5) as obtained from figure 2 is presented in figure 4. As can be seen from figure 2, only one neuron is needed per equation. The neurons are in turn realized by employing only two multi-output CDTAs per neuron.
Fig. 4 The proposed circuit applied to a 2 – variable problem (5)
The circuit of figure 4 was simulated using PSPICE simulation tool. CDTAs were realized using the bipolar junction transistor implementation [5] shown in figure 3. Not explicitly shown in figure 4 is the output port ‘Z’ of the CDTAs. A fixed valued resistor (RZ = 10KΩ) was connected between the Z-output and ground for each of the CDTAs to set the value of the current gain at a high value. The resultant transfer characteristics of the CDTA are presented in figure 5
from which it can be readily verified that the tanh(.) function is indeed obtained as mentioned in (3). For the purpose of PSPICE simulations, the supply voltages were set to VCC = 3V, VEE = -3V and all the biasing currents of the CDTA were kept at 1000µA. Further, to get the output currents according to (5) the emitter areas of transistors in the output stage of the CDTAs were set to provide the required current scaling. The corresponding scaling factors are denoted on the CDTA symbol as ‘X’, and ‘2X’ where ‘2X’ means a current output double that of the ‘X’ output. Routine mathematical analysis of (5) gives the solution as I1 = 15 and I2 = 10. The results of PSPICE simulation of the circuit of fig. 2, shown in figure 6, are found to be I1 = 14.817 µA and I2 = 9.872 µA which are in close agreement with the algebraic solution. Further, another noteworthy point is that the solution time of the proposed circuit is of the order of nanoseconds thereby making it suitable for real-time signal processing applications.
Fig. 5 Transfer characteristics of the CDTA based current-mode comparator
286
IMPACT-2009
Fig. 6 Simulation results for the chosen 2 – variable problem (5)
V. CONCLUSIONS
In this paper we have presented a novel current-mode approach to solve simultaneous linear equations in n variables, which uses n-neurons and n-synapses. Each neuron requires one current-mode amplifier and each synapse is implemented using one current-mode comparator. Both the amplifier and the comparator were implemented using CDTAs. No resistances were needed to realize the weighted synaptic interconnections. This resulted in significant reduction in hardware over the existing schemes [2, 3, and 4].
The proposed network was tested on a sample problem set of 2 simultaneous linear equations and the simulation results confirm the validity of the approach. The time to arrive at the solution was of the order of nanoseconds.
REFERENCES [1] D. Jiang, “Analog Computing for Real-Time solution of Time-Varying
Linear Equations,” Proc. ICCCAS, Volume 2, 27-29 June 2004. [2] Y. Xia, J. Wang and D. L. Hung, “Recurrent Neural Networks for
Solving Linear Inequalities and Equations,” IEEE Tran. Cir. & Sys - I, vol. 46, no. 4, Apr 1999.
[3] A. Cichocki and R. Unbehauen, “Neural Networks for Solving Systems of Linear Equations and Related Problems,” IEEE Tran. Cir. & Sys - I, vol. 39, no. 2, Feb 1992.
[4] J. Wang, “Electronic Realization of Recurrent Neural Network for Solving Simultaneous Linear Equations,” Electronics Letters, vol. 28, no. 5, Feb 1992.
[5] N. A. Shah, M. Quadri and S. Z. Iqbal, “CDTA based universal transadmittance filter,” Analog Integr Circ Sig Process, vol. 52, Page(s): 65–69, Sept. 2007.
[6] S. A. Rahman and M. S. Ansari, “Solution of Simultaneous Linear Equations using Feedback Neural Network,” Proc. RASIET, IET, Alwar, India, Dec 2007.
[7] S. A. Rahman, Jayadeva and S. C. Duttaroy, “Neural network approach to graph colouring,” Electronics Letters, vol. 35, no. 14, July 1999
[8] L. Song, M. I. Elmasry and A. Vannelli, “Analog Neural Network Building Blocks Based on Current Mode Subthreshold Operation,” Proc. ISCAS’93, Page(s):2462 - 2465, 3-6 May 1993.
[9] R. W. Newcomb and J. D. Lohn, “Analog VLSI for neural networks,” in The Handbook of Brain Theory and Neural Networks. Cambridge, MA: The MIT Press, 1998.
[10] Web Link: http://en.wikipedia.org/wiki/Simultaneous_linear_equations [11] J. L. Gustafson, “The Quest for Linear Equation Solvers and the
Invention of Electronic Digital Computing,” Proc. JVA’06, IEEE John Vincent Atanasoff 2006 International Symposium on Modern Computing, 2006
[12] K. Zhang, G. Ganis and M. I. Sereno, “Anti-Hebbian synapses as a linear equation solver,” Proc. International Conference on Neural Networks, 9 – 12 June 1997, vol. 1, Page(s): 387 - 389
287
International MultiConference on Intelligent Systems and Nanotechnology (IISN-2010), February 26-28, 2010 Computational Intelligence Laboratory, Institute of Science and Technology, Klawad-133105, Haryana, India
QPP: Solution using feedback neural network
employing transconducting synaptic interconnections
Mohd. Samar Ansari and Syed Atiqur Rahman
Department of Electronics Engineering, Aligarh Muslim University, Aligarh, India
(Emails: [email protected], [email protected])
Abstract—This paper presents a neural circuit for solving
quadratic programming problems. The objective is to minimize
a quadratic cost function subject to linear constraints. The
proposed circuit employs non-linear feedback, in the form of
unipolar comparators, to introduce transcendental terms in the
energy function ensuring fast convergence to the solution.
Further, the use of resistors to generate weighted inputs to the
neurons is avoided. Instead, transconducting elements are
utilized to directly generate the required scaled currents.
PSPICE simulation results are presented for a chosen
optimization problem and are found to agree with the algebraic
solution.
Index Terms—Neural network applications, Neural network
hardware, Nonlinear circuits, Quadratic Programming,
Optimisation
I. INTRODUCTION
n the discipline of constrained optimization, problems with
nonlinear objective functions are usually approximated by
a second-order (quadratic) system and solved approximately
by a standard quadratic programming technique. Traditional
methods for solving quadratic programming problems
typically involve an iterative process, but long computational
time limits their usage. There is an alternative approach to
solution of this problem. It is to exploit the artificial neural
networks (ANN's) which can be considered as an analog
computer relying on a highly simplified model of neurons [1].
ANN's have been applied to several classes of constrained
optimization problems and have shown promise for solving
such problems more effectively. For example, the Hopfield
neural network has proven to be a powerful tool for solving
some of the optimization problems. Tank and Hopfield
first proposed a neural network for solving mathematical
programming problems, where a linear programming
problem was mapped into a closed-loop network [2]. Later,
the dynamical approach was extended for solving quadratic
programming problems. Over the past two decades several
neural-network architectures for solving quadratic
programming problems have been proposed by Kennedy and
Chua [3], Maa and Shanblatt [4], Chen and Feng [5], Wu et
al. [6], Xia [7]. These networks depend on the network
parameters or use expensive analog multipliers. In this paper,
we present a hardware solution to the problem of solving a
quadratic programming problem subject to linear constraints.
The paper is organized as follows. Section II describes the
proposed neural network to minimize a quadratic polynomial
in 2 variables subject to 2 linear constraints. Section III
contains results of PSPICE simulation. Some concluding
remarks appear in section IV.
II. PROPOSED NETWORK
Let the function to be minimized be
𝐹 =1
2𝐶11𝑉1
2 + 𝐶12𝑉1𝑉2 +1
2𝐶22𝑉2
2 (1)
Subject to
𝑎11𝑉1 + 𝑎12𝑉2 ≤ 𝑏1 (2)
𝑎21𝑉1 + 𝑎22𝑉2 ≤ 𝑏2 (3)
Where V1, and V2 are the variables and aij, bi and cij are
constants. The proposed neural-network based circuit to
minimize (1) subject to (2, 3) is presented in fig. 1.
Vi
Vout
+
_
Vi
Vout
+
_
b1
b2
x1
x2
G1
G2
a11V1+ a12V2
a21V1+ a22V2
gc11 x1
gc12 x2
R11
R21
+
_ V1
Rp1 C1
+_
V2
RpiC2
R22
R12
gc22 x2
gc21 x1
A
B
Fig. 1 The proposed neural network for solving a quadratic programming problem in 2 variables subject to 2 linear constraints
As can be seen from fig. 1, individual equations from the set
of constraints are passed through non-linear synapses which
are realized using voltage-mode unipolar comparators
followed by multi-output transconducting cells. From fig. 1,
the unipolar comparator outputs can be modeled as
𝑥1 =𝑉𝑚
2 𝑡𝑎𝑛ℎ 𝛽 𝑎11𝑉1 + 𝑎12𝑉2 + 1 (4)
𝑥2 =𝑉𝑚
2 𝑡𝑎𝑛ℎ 𝛽 𝑎21𝑉1 + 𝑎22𝑉2 + 1 (5)
where 𝛽 is the open-loop gain of the comparator (practically
very high), ±Vm are the output voltage levels of the
I
International MultiConference on Intelligent Systems and Nanotechnology (IISN-2010), February 26-28, 2010 Computational Intelligence Laboratory, Institute of Science and Technology, Klawad-133105, Haryana, India
comparator and V1 and V2 are the neuron outputs. The outputs
of the comparators are fed to neurons having weighted inputs.
The neurons are realized by using opamps and the weights
are implemented using transconductance amplifiers. The
currents arriving to the neuron from various synapses get
added up at the input of the neuron. Rpi and Cpi are the input
resistance and capacitance of the opamp corresponding to the
i-th neuron. These parasitic components are included to
model the dynamic nature of the opamp.
Node equation for node ‘A’ gives the equation of motion of
the first neuron in the state space as
𝐶1𝑑𝑢1
𝑑𝑡= 𝑥1𝑔𝑐11 + 𝑥2𝑔𝑐21 +
𝑉1
𝑅11+
𝑉2
𝑅21 −
𝑢1 𝑔𝑐11 + 𝑔𝑐21 +1
𝑅11+
1
𝑅21+
1
𝑅𝑝1 (6)
On simplification, (3) yields
𝐶1𝑑𝑢1
𝑑𝑡= 𝑥1𝑔𝑐11 + 𝑥2𝑔𝑐21 +
𝑉1
𝑅11+
𝑉2
𝑅21 − 𝑢1
1
𝑅1
(7)
Where
1
𝑅1=
1
𝑅11+
1
𝑅21+ 𝑔𝑐11 + 𝑔𝑐21 +
1
𝑅𝑝1 (8)
Similarly, for node ‘B’ we can write
𝐶2𝑑𝑢2
𝑑𝑡= 𝑥1𝑔𝑐12 + 𝑥2𝑔𝑐22 +
𝑉1
𝑅12+
𝑉2
𝑅22 − 𝑢2
1
𝑅2
(9)
Where
1
𝑅2=
1
𝑅12+
1
𝑅22+ 𝑔𝑐12 + 𝑔𝑐22 +
1
𝑅𝑝2 (10)
Where ui is the internal state of the i-th neuron and gcji is the
voltage-to-current conversion factor of the j-th
transconductance block for the i-th output current. As is
shown later in this section, these weights are governed by the
coefficients of (2, 3). Moreover, it can be shown that the
network of fig. 2 may be associated with an ‘energy’ function
of the form
𝐸 =1
2 𝐶11𝑉1
2 + 2𝐶12𝑉1𝑉2 + 𝐶22𝑉22
+𝑉𝑚
𝛽 ln cosh 𝛽 𝑎11𝑉1 + 𝑎12𝑉2– 𝑏1
+𝑉𝑚
𝛽 ln cosh 𝛽 𝑎21𝑉1 + 𝑎22𝑉2– 𝑏2
+𝑉𝑚
2 𝑎11𝑉1 + 𝑎12𝑉2 + 𝑎21𝑉1 + 𝑎22𝑉2 (11)
From (11), it follows that
𝜕𝐸
𝜕𝑉1= 𝑉𝑚 𝑎11 tanh 𝛽(𝑎11𝑉1 + 𝑎12𝑉2 − 𝑏1) +
𝑉𝑚𝑎21tanh 𝛽(𝑎21𝑉1 + 𝑎22𝑉2 − 𝑏2) − 𝑢1 1
𝑅1 (12)
Also, if ‘E’ is the Energy Function, it must satisfy the
following condition [8].
𝜕𝐸
𝜕𝑉1= 𝐾 𝐶1
𝑑𝑢1
𝑑𝑡 (13)
Where ‘K’ is a constant of proportionality and has the
dimensions of resistance. Comparing (7) and (12) according
to (13) using (4) and (5) yields
𝑔11 = 𝑎11
𝐾 , 𝑔21 = 𝑎21
𝐾 (14)
𝑅11 = 𝐾 𝑐 11, 𝑅21 = 𝐾 𝑐 21
(15)
The values of the transconductance factors (gji) of Fig. 1
can be easily calculated by choosing a suitable value of ‘K’
and then using (14). Analysis on similar lines can be
performed to obtain the values of the synaptic weights for the
second neuron and are presented in (16) and (17) below.
𝑔12 = 𝑎12
𝐾 , 𝑔22 = 𝑎22
𝐾 (16)
𝑅12 = 𝐾 𝑐 12, 𝑅22 = 𝐾 𝑐 22
(17)
III. SIMULATION RESULTS
This section deals with the application of the proposed
network to task of minimizing the objective function
𝑉12 + 4𝑉1𝑉2 + 5𝑉2
2 (18)
subject to
𝑉1 − 𝑉2 ≤ −1
𝑉1 + 𝑉2 ≤ 1
The values of resistances acting as the weights on the neurons
are obtained from (15). For the purpose of simulation, the
value of ‘K’ was chosen to be 1 KΩ. Using K = 1 KΩ in (14,
15) gives
gc11 = gc12 = gc21 = gc22 = 1/K = 1 R11 = 1.66 KΩ, R21 = 1 KΩ, R12 = 1.66 KΩ, R22 = 1 KΩ
For the purpose of PSPICE simulations, the voltage
comparator and the multi-output transconductance amplifier
for each neuron were realized as a single block using a
slightly modified circuit of the differential-input, high-gain,
active-loaded transconductance amplifier [9] shown in Fig. 2.
Standard BSIM3 0.35µm parameters were used for the
purpose of simulations. The supply voltages were set to VDD
= 15V, VSS = 0V and VBB = 1V. The aspect ratios of the
NMOS and PMOS transistors were taken to be 1.4µm/1.4µm
and 2.8µm/1.4µm respectively. Further, to get the output
currents according to (14, 16) the W/L ratios of transistors
M6, M7, M8 and M9 were set to provide the required current
scaling. Choosing K=1 and using (14) results in Iout1 = Iout2 for
G1 block and Iout1 = Iout2 for the G2 block.
(19)
International MultiConference on Intelligent Systems and Nanotechnology (IISN-2010), February 26-28, 2010 Computational Intelligence Laboratory, Institute of Science and Technology, Klawad-133105, Haryana, India
M2
M1
M3
M4
M5
M6
M7
VDD
VSS
VBB
V2
V1 I
out1
M8
M9
Iout2
Fig. 2 CMOS implementation of the comparator and transconductance
blocks
Fig. 3 Simulation results for the proposed circuit
Routine mathematical analysis of (18, 19) yields: 𝑉1= -0.584,
𝑉2= 0.416. The results of PSPICE simulation are presented in
fig. 3. From the plots of the neuron output voltages, it can be
seen that V(1) = -0.58 V and V(2) = 0.41 V which are very
near to the algebraic solution thereby confirming the validity
of the approach.
IV. CONCLUSION
In this paper we have described a CMOS compatible
approach to solve a quadratic programming problem in 2
variables subject to 2 linear constraints, which uses 2-neurons
and 2-synapses. Each neuron requires one opamp and each
synapse is implemented using a differential transconductance
element. This results in significant reduction in hardware
over the existing schemes [3-8]. The proposed network was
tested on a sample problem of minimizing a quadratic
function in 2 variables and the simulation results confirm the
validity of the approach.
REFERENCES
[1] J.J. Hopfield, D.W. Tank, (1985) ‘Neural’ computation of decisions
optimization problems, Biological Cybern., 52, 141-152.
[2] D.W. Tank, J.J. Hopfield, (1986) Simple neural optimization networks: an A/D converter, signal decision network, and linear programming
circuit, IEEE Trans. Circ. Syst. CAS, 33, 533–541.
[3] M.P. Kennedy, L.O. Chua, (1988) Neural networks for nonlinear programming, IEEE Trans. Circ. Syst., 35, 554–562.
[4] C.Y. Maa, M.A. Shanblatt, (1992) A two-phase optimization neural
network, IEEE Trans. Neural Network 3 (6), 580 – 594. [5] Y.-H. Chen, S.-C. Fang, (1998) Solving convex programming problems
with equality constraints by neural networks, Comput. Math. Appl., 36
(7), 41–68. [6] X.Y. Wu, Y.S. Xia, J. Li, W.K. Chen, (1996) A high performance
neural network for solving linear and quadratic programming problems,
IEEE Trans. Neural Networks, 7 (3), 643–651. [7] Y. Xia, (1996) A new neural network for solving linear and quadratic
programming problems, IEEE Trans. Neural Networks 7 (6), 1544–
1547. [8] S. A. Rahman, Jayadeva, S. C. Dutta Roy, (1999) Neural network
approach to graph colouring, Electronics Letters, 35, 1173 – 1175.
[9] R. W. Newcomb, J. D. Lohn, (1998) Analog VLSI for neural networks, The Handbook of Brain Theory and Neural Networks. Cambridge,
MA: The MIT Press.
A Voltage-Mode Analog Circuit for Solving Linear Programming Problems
Mohd. Samar Ansari
Department of Electronics Engineering
Aligarh Muslim University
Aligarh, India email: [email protected]
Syed Atiqur Rahman
Department of Electronics Engineering
Aligarh Muslim University
Aligarh, India email: [email protected]
Abstract—This paper presents a neural circuit for solving linear
programming problems. The objective is to minimize a first order
cost function subject to linear constraints. The dynamic analog
circuit, consisting of N identical units for N variable problem, can
solve the general LPP and always converges to the optimal
solution in constant time, irrespective of the initial conditions,
which is of the order of its time constant. The proposed circuit
employs non-linear feedback, in the form of unipolar comparators,
to introduce transcendental terms in the energy function ensuring
fast convergence to the solution. Further, the use of resistors to
generate weighted inputs to the neurons is avoided. Instead,
transconducting elements are utilized to directly generate the
required scaled currents. PSPICE simulation results are presented
for a chosen optimization problem and are found to agree with the
algebraic solution.
Index Terms—Neural network applications, Neural network
hardware, Nonlinear circuits, Linear Programming,
Optimization
I. INTRODUCTION
inear programming is a considerable field of optimization for
several reasons. Many practical problems in operations
research can be expressed as linear programming problems.
Certain special cases of linear programming, such as network flow
problems and multicommodity flow problems are considered
important enough to have generated much research on specialized
algorithms for their solution [1-6]. A number of algorithms for
other types of optimization problems work by solving LP
problems as sub-problems. In an LPP problem, a linear function is
optimized subject to certain linear constraints. The general LPP
problem can be stated as:
Minimize 𝑷𝑇𝑽 (1)
Subject to 𝒂𝑽 ≤ 𝒃 (2)
where 𝑽RN is a column vector of decision variables, 𝑷RN is a
column vector of cost coefficient, 𝒃RM is a column vector of
constant bounds, 𝒂RMxN is a constraint coefficient matrix, M
N, and the superscript T denotes the transpose operator. The M
constraints specified by (2) define the feasible region in the N
dimensional Euclidean space in which the cost function is required
to be minimized [2].
Traditional methods for solving linear programming problems
typically involve an iterative process, but long computational time
limits their usage. There is an alternative approach to solution of
this problem. It is to exploit the artificial neural networks (ANN's)
which can be considered as an analog computer relying on a
highly simplified model of neurons [1]. ANN's have been applied
to several classes of constrained optimization problems and have
shown promise for solving such problems more effectively. For
example, the Hopfield neural network has proven to be a powerful
tool for solving some of the optimization problems. Tank and
Hopfield first proposed a neural network for solving
mathematical programming problems, where a linear
programming problem was mapped into a closed-loop network
[2]. Over the past two decades several neural-network
architectures for solving linear programming problems have been
proposed by Wang [3], Xia [4], Xia, Wang & Hung [5] and Malek
and Yari [6]. Wang‘s realization of a neural network for solving
linear programming problems utilizes 3n amplifiers, n2+3n
resistances and n capacitors. The circuit of [4] requires 2m2+4mn
amplifiers, 2m2+4mn+3m+3 summers, n+m integrators, and n
simple limiters. The circuit of [5] consists of m2+2mn amplifiers,
3n summers, n integrators, and n simple limiters. And although no
circuit implementation of the proposed scheme is given in [6], it is
evident that it would not be simple to actually realize in hardware.
Therefore, the search still continues for a working circuit, to solve
LPP, which is feasible to implement in integrated forms and is fast
enough for utilization in real-time applications where the time-to-
solve is needed to be of the order of tenths of milliseconds [5].
In this paper, a hardware solution to the problem of solving a
linear programming problem subject to linear inequality
constraints is presented. The proposed circuit employs only n
amplifiers, m comparators and mn+m+n resistances. It is evident
that the circuit complexity of the proposed scheme is much
reduced as compared to existing ones [3-6].
The paper is organized as follows. Section II describes the
proposed neural network to minimize a first order polynomial in 2
variables subject to 2 linear constraints. Section III contains results
of PSPICE simulation. Some concluding remarks appear in
section IV.
II. PROPOSED NETWORK
Let the function to be minimized be
𝐹 = 𝑃1𝑉1 + 𝑃2𝑉2 (3) Subject to
𝑎11𝑉1 + 𝑎12𝑉2 ≤ 𝑏1 (4)
𝑎21𝑉1 + 𝑎22𝑉2 ≤ 𝑏2 (5)
Where V1, and V2 are the variables and aij, bi and cij are constants.
The proposed neural-network based circuit to minimize (3) subject
to (4, 5) is presented in fig. 1.
L
Vi
Vout
+
_
Vi
Vout
+
_
b1
b2
x1
x2
G1
G2
a11V1+ a12V2
a21V1+ a22V2
gc11 x1
gc12 x2
R1
+
_
V1Rp1 C1
+_
V2RpiC2
R2
gc22 x2
gc21 x1
A
B
P1
P2
Fig. 1 The proposed neural network for solving a linear programming
problem in 2 variables subject to 2 linear constraints As can be seen from fig. 1, individual equations from the set of
constraints are passed through non-linear synapses which are
realized using voltage-mode unipolar comparators followed by
multi-output transconducting cells. From fig. 1, the unipolar
comparator outputs can be modeled as
𝑥1 =𝑉𝑚
2 𝑡𝑎𝑛ℎ 𝛽 𝑎11𝑉1 + 𝑎12𝑉2 + 1 (6)
𝑥2 =𝑉𝑚
2 𝑡𝑎𝑛ℎ 𝛽 𝑎21𝑉1 + 𝑎22𝑉2 + 1 (7)
where 𝛽 is the open-loop gain of the comparator (practically very
high), ±Vm are the output voltage levels of the comparator and V1
and V2 are the neuron outputs. The outputs of the comparators are
fed to neurons having weighted inputs. The neurons are realized
by using opamps and the weights are implemented using
transconductance amplifiers. The currents arriving to the neuron
from various synapses get added up at the input of the neuron. Rpi
and Cpi are the input resistance and capacitance of the opamp
corresponding to the i-th neuron. These parasitic components are
included to model the dynamic nature of the opamp.
Node equation for node ‗A‘ gives the equation of motion of the
first neuron in the state space as
𝐶1𝑑𝑢1
𝑑𝑡= 𝑥1𝑔𝑐11 + 𝑥2𝑔𝑐21 +
𝑃1
𝑅1 − 𝑢1 𝑔𝑐11 +
𝑔𝑐21 +1
𝑅1+
1
𝑅𝑝1 (8)
On simplification, (3) yields
𝐶1𝑑𝑢1
𝑑𝑡= 𝑥1𝑔𝑐11 + 𝑥2𝑔𝑐21 +
𝑃1
𝑅1 − 𝑢1
1
𝑅𝑒𝑞 1 (9)
Where
1
𝑅𝑒𝑞 1=
1
𝑅1+ 𝑔𝑐11 + 𝑔𝑐21 +
1
𝑅𝑝1 (10)
Similarly, for node ‗B‘ we can write
𝐶2𝑑𝑢2
𝑑𝑡= 𝑥1𝑔𝑐12 + 𝑥2𝑔𝑐22 +
𝑃2
𝑅2 − 𝑢2
1
𝑅𝑒𝑞 2 (11)
Where
1
𝑅𝑒𝑞 2=
1
𝑅2+ 𝑔𝑐12 + 𝑔𝑐22 +
1
𝑅𝑝2 (12)
Where ui is the internal state of the i-th neuron and gcji is the
voltage-to-current conversion factor of the j-th transconductance
block for the i-th output current. As is shown later in this section,
these weights (gcji) are governed by the coefficients of (4, 5).
Moreover, it can be shown that the network of fig. 1 may be
associated with an ‗energy‘ function of the form
𝐸 = 𝑃1𝑉1 + 𝑃2𝑉2 + 𝑉𝑚
𝛽 ln cosh 𝛽 𝑎11𝑉1 +
𝑎12𝑉2– 𝑏1 +𝑉𝑚
𝛽 ln cosh 𝛽 𝑎21𝑉1 + 𝑎22𝑉2– 𝑏2 +
𝑉𝑚
2 𝑎11𝑉1 + 𝑎12𝑉2 + 𝑎21𝑉1 + 𝑎22𝑉2 (13)
From (11), it follows that
𝜕𝐸
𝜕𝑉1= 𝑃1 + 𝑉𝑚 𝑎11tanh 𝛽(𝑎11𝑉1 + 𝑎12𝑉2 − 𝑏1) +
𝑉𝑚𝑎21tanh 𝛽(𝑎21𝑉1 + 𝑎22𝑉2 − 𝑏2) (14)
Also, if ‗E‘ is the Energy Function, it must satisfy the following
condition [7].
𝜕𝐸
𝜕𝑉1= 𝐾 𝐶1
𝑑𝑢1
𝑑𝑡 (15)
Where ‗K‘ is a constant of proportionality and has the dimensions
of resistance. Comparing (9) and (14) according to (15) using (6)
yields
𝑔𝑐11 = 𝑎11
𝐾 , 𝑔𝑐21 = 𝑎21
𝐾 (16)
𝑅1 = 𝐾 (17)
The values of the transconductance factors (gji) of Fig. 1 can be
easily calculated by choosing a suitable value of ‗K‘ and then
using (16). Analysis on similar lines can be performed to obtain
the values of the synaptic weights for the second neuron and are
presented in (18) and (19) below.
𝑔𝑐12 = 𝑎12
𝐾 , 𝑔𝑐22 = 𝑎22
𝐾 (18)
𝑅2 = 𝐾 (19)
III. SIMULATION RESULTS
This section deals with the application of the proposed network to
task of minimizing the objective function
5𝑉1 + 2𝑉2 (20) subject to
𝑉1 + 𝑉2 ≥ 1
𝑉2 ≤ 1
The values of resistances acting as the weights on the neurons are
obtained from (17, 19). For the purpose of simulation, the value of
‗K‘ was chosen to be 1 KΩ. Using K = 1 KΩ in (16 through 19)
gives
gc11 = gc21 = gc22 = 1/K = 1; gc12 = 0 R1 = 1 KΩ, R2 = 1 KΩ
For the purpose of PSPICE simulations, the voltage comparator
and the multi-output transconductance amplifier for each neuron
were realized as a single block using a slightly modified circuit of
(21)
(22)
the differential-input, high-gain, active-loaded transconductance
amplifier [8] shown in Fig. 2. Standard BSIM3 0.35µm parameters
were used for the purpose of simulations. The supply voltages
were set to VDD = 15V, VSS = 0V and VBB = 1V. The aspect
ratios of the NMOS and PMOS transistors were taken to be
1.4µm/1.4µm and 2.8µm/1.4µm respectively. Further, to get the
output currents according to (22) the W/L ratios of transistors M6,
M7, M8 and M9 were set to provide the required current scaling.
M2
M1
M3
M4
M5
M6
M7
VDD
VSS
VBB
V2
V1 I
out1
M8
M9
Iout2
Fig. 2 CMOS implementation of the comparator and transconductance
blocks
Routine mathematical analysis of (20, 21) yields: 𝑉1= 0, 𝑉2= 1.
The results of PSPICE simulation are presented in fig. 3. From the
plots of the neuron output voltages, it can be seen that V(1) = 0 V
and V(2) = 1 V which correspond exactly to the algebraic solution
thereby confirming the validity of the approach.
Fig. 3 Simulation results for the proposed circuit
IV. CONCLUSION
In this paper, a CMOS compatible approach to solve a linear
programming problem in 2 variables subject to 2 linear
constraints, which uses 2-neurons and 2-synapses, is described.
Each neuron requires one opamp and each synapse is implemented
using a differential transconductance element. This results in
significant reduction in hardware over the existing schemes [3-6].
The proposed network was tested on a sample problem of
minimizing a quadratic function in 2 variables and the simulation
results confirm the validity of the approach.
REFERENCES
[1] J.J. Hopfield, D.W. Tank, ―‗Neural‘ computation of decisions
optimization problems,‖ Biological Cybern., 52, 141-152, 1985
[2] D. W. Tank and J. J. Hopfield, ―Simple Neural Optimization Networks: An A/D Converter, Signal Decision Circuit, and A Linear
Programming Circuit,” IEEE Trans. Circuits and Systems, Vol.
CAS-33, No.5, pp. 533-541, May 1986. [3] J. Wang, ―Analysis and Design of a Recurrent Neural Network for
Linear Programming‖, IEEE Trans. Circuits and Systems–1:
Fundamental Theory and Applications, Vol.40, No.9, pp. 613-618, Sept.1993.
[4] Y. Xia, ―A New Neural Network for Solving Linear and Quadratic
Programming Problems,‖ IEEE Trans. Neural Networks, Vol.7, No.6, pp. 1544-1547, Nov 1996.
[5] Y. Xia, J. Wang and D. L. Hung, ―Recurrent Neural Networks for
Solving Linear Inequalities and Equations,‖ IEEE Trans. Circuits and Systems–1: Fundamental Theory and Applications, Vol.46,
No.4, pp. 452-462, April 1999.
[6] A. Malek and A. Yari., ―Primal–dual solution for the linear programming problems using neural networks,‖ Applied
Mathematics and Computation, 167, 198–211, 2005.
[7] S. A. Rahman, Jayadeva, S. C. Dutta Roy, ―Neural network approach to graph colouring,‖ Electronics Letters, 35, 1173 – 1175,
1999
[8] M. S. Ansari and S. A. Rahman, ―A non-linear neural circuit for solving system of simultaneous linear equations,‖ Proc. IMPACT-
2009, AMU, Aligarh, March 2009.
CERA09
DVCC-Based Non-Linear Feedback Neural Circuit
for Solving Quadratic Programming Problems
Mohd. Samar Ansari and Syed Atiqur Rahman
1,2Department of Electronics Engineering, Aligarh Muslim University, Aligarh, India
(Emails:[email protected], [email protected]
2)
Abstract— This paper presents a non-linear feedback neural
network for solving quadratic programming problems. The
objective is to minimize a quadratic cost function subject to
linear constraints. The proposed circuit employs non-linear
feedback, in the form of unipolar comparators realized using
DVCCs and diodes, to introduce transcendental terms in the
energy function ensuring fast convergence to the solution.
PSPICE simulation results are presented for a chosen
optimization problem and are found to agree with the algebraic
solution.
Index Terms— Neural Networks, Non – Linear circuits,
Feedback Neural Networks, Quadratic Programming,
Dynamical Systems
I. INTRODUCTION
uadratic Programming problems are very important in the
field of optimization. They arise in many applications
such as constrained least mean square estimation and the
classical newsvendor problem. Besides its wide applications,
quadratic programming is also of theoretic meaning, because
it forms a basis for solving some general nonlinear
programming problems. In the discipline of constrained
optimization, problems with nonlinear objective functions are
usually approximated by a second-order (quadratic) system
and solved approximately by a standard quadratic
programming technique. Traditional methods for solving
quadratic programming problems typically involve an
iterative process, but long computational time limits their
usage. There is an alternative approach to solution of this
problem. It is to exploit the artificial neural networks
(ANN's) which can be considered as an analog computer
relying on a highly simplified model of neurons [1]. ANN's
have been applied to several classes of constrained
optimization problems and have shown promise for solving
such problems more effectively. For example, the Hopfield
neural network has proven to be a powerful tool for solving
some of the optimization problems. Tank and Hopfield
first proposed a neural network for solving mathematical
programming problems, where a linear programming
problem was mapped into a closed-loop network [1]. Later,
the dynamical approach was extended for solving quadratic
programming problems. Over the past two decades several
neural-network architectures for solving quadratic
programming problems have been proposed by Kennedy and
Chua [2], Maa and Shanblatt [3], Wu et al. [4], Xia [5].
These networks depend on the network parameters or use
expensive analog multipliers. More recently, Malek and
Alipour [6] proposed a recurrent neural network that is able
to solve quadratic programming problems. This architecture
is advantageous in the sense that there is no need to set
network parameters. Moreover, the number of analog
multipliers is reduced. This reduces the hardware complexity
to a certain extent. However, the network still has a complex
hardware in the sense that the main processing element
(alpha) would be cumbersome to implement in hardware. In
this paper, we present a hardware solution to the problem of
solving a quadratic programming problem subject to linear
constraints. The proposed architecture uses non-linear
feedback which leads to a new energy function that involves
transcendental terms. This transcendental energy function is
fundamentally different from the standard quadratic form
associated with Hopfield network and its variants. The
hardware complexity of the proposed circuit compares
favourably with the existing hardware implementations.
The remainder of this paper is arranged as follows:
Section–II outlines the basic problem and the mathematical
formulation on which the development of the proposed
network will be based. Section–III contains the circuit
implementation of the proposed network for a set of sample
problem in two variables. SPICE simulation results of the
proposed circuit are also presented. Issues that are expected
to arise in actual monolithic implementations are discussed in
Section–IV. Concluding remarks are presented in Section–V.
II. PROPOSED NEURAL NETWORK
Let the function to be minimized be
𝐹 =1
2𝐶11𝑉1
2 + 𝐶12𝑉1𝑉2 +1
2𝐶22𝑉2
2 (1)
Subject to
𝑎11𝑉1 + 𝑎12𝑉2 ≤ 𝑏1 (2)
𝑎21𝑉1 + 𝑎22𝑉2 ≤ 𝑏2 (3)
Where V1, and V2 are the variables and aij, bi and cij are
constants. The proposed neural-network based circuit to
minimize (1) subject to (2, 3) is presented in fig. 1.
Q
CERA09
b1
x1s11Ix
Z+
1
XY2
Y1
DVCC-1
a11V1+ a12V2
Z+
2
s12Ix
b2
x1s21Ix
Z+
1
XY2
Y1
DVCC-2
a21V1+ a22V2
Z+
2
s22Ix
+
_V1
+_
Rp2Cp2
V2
Rp1 Cp1
R11
R22
R21
R12
A
B
Fig. 1 The proposed feedback neural circuit to solve a quadratic
programming problem in 2-variables with 2 linear constraints
As can be seen from fig. 1, individual equations from the
set of constraints are passed through non-linear synapses
which are realized using DVCC-based unipolar comparators.
x1
Z+
X
Y2
Y1
DVCC
Z-
VY1
VY2
VX
IX
IZ+
IZ-
Fig. 2 Symbolic Representation of DVCC
The symbolic diagram of the DVCC is shown in fig. 2. The
device can be characterized by the following port relations:
𝑉𝑋
𝐼𝑌1
𝐼𝑌2
𝐼𝑍+
𝐼𝑍−
=
0001
−1
10000
−10000
00000
00000
𝐼𝑋𝑉𝑌1
𝑉𝑌2
𝑉𝑍+
𝑉𝑍−
(4)
Although both Z+ and Z- types of current outputs are
mentioned in (4), the DVCCs used in the proposed network
use only Z+ type of outputs. For the DVCC, comparator
action can be achieved by putting 𝑅𝑋 → 0 i.e. by directly
grounding the X-terminal [7]. For such a case, the current in
the X-port will saturate and can be written as
𝐼𝑋 = 𝐼𝑚 tanh 𝛽 𝑉𝑌1– 𝑉𝑌2 (5)
Where β is the open-loop gain of the comparator (practically
very high) and ±𝐼𝑚 are the saturated output current levels of
the comparator. Equation (5) can also be written in an
equivalent notation as
𝐼𝑋 = 𝑔𝑉𝑚 tanh 𝛽 𝑉𝑌1– 𝑉𝑌2 (6)
Where 𝑔 is the transconductance from the input voltage ports
(Y1 and Y2) to the output current port (X) and ±𝑉𝑚 are the
biasing voltages of the DVCC-based comparator.
By virtue of DVCC action, this current will be transferred to
the Z+ ports as
𝐼𝑧+ = 𝑠𝐼𝑋 = 𝑠𝑔𝑉𝑚 tanh 𝛽 𝑉𝑌1– 𝑉𝑌2 (7)
Where 𝑠 is the current scaling factor introduced during the
current conveying process from the X-port to the Z+ port.
Ideally, 𝑠 = 1. However, by altering the aspect ratios of the
transistors used in the output stages of the DVCC, 𝐼𝑧+ can be
made a scaled replica of 𝐼𝑋 . In that case, the value of 𝑠 will be
different from unity. This property will be utilized in the
design of the multi-output DVCCs needed for the proposed
circuit.
The outputs of the comparators are fed to neurons having
weighted inputs. These weighted neurons are realized by
using opamps where the scaled currents coming from various
comparators act as weights. Weighing is also provided by the
self-feedback resistances (R11 & R22 for neurons 1 & 2
respectively) and the cross-feedback resistances (R21 and R12
for neurons 1 & 2 respectively). Rpi and Cpi are the input
resistance and capacitance of the opamp corresponding to the
i-th neuron. These parasitic components are included to
model the dynamic nature of the opamp.
Node equation for nodes ‘A’ and ‘B’ give the equations of
motions of the first and second neurons as
𝐶𝑝1𝑑𝑢1
𝑑𝑡= 𝑠11𝐼′𝑋1 + 𝑠21𝐼′𝑋2 +
𝑉1
𝑅11+
𝑉2
𝑅21 − 𝑢1
1
𝑅1 (8)
Where, 𝑠11𝐼′𝑋1 and 𝑠21𝐼′𝑋2 represent the currents 𝑠11𝐼𝑋1 and
𝑠21𝐼𝑋2after passing through diodes as shown in Fig. 1 and
1
𝑅1=
1
𝑅𝑝1+
1
𝑅11+
1
𝑅21 (9)
And,
𝐶𝑝2𝑑𝑢2
𝑑𝑡= 𝑠12𝐼′𝑋1 + 𝑠22𝐼′𝑋2 +
𝑉1
𝑅12+
𝑉2
𝑅22 − 𝑢2
1
𝑅2 (10)
Where, 𝑠12𝐼′𝑋1 and 𝑠22𝐼′𝑋2 represent the currents 𝑠12𝐼𝑋1 and
𝑠22𝐼𝑋2 after passing through diodes as shown in Fig. 1 and
1
𝑅2=
1
𝑅𝑝2+
1
𝑅12+
1
𝑅22 (11)
In equations (8) and (10) ui is the internal state of the i-th
neuron and 𝑠𝑗𝑖 is the current scaling factor at the j-th output of
i-th DVCC. As is shown later in this section, these weights
are governed by the constraint inequalities (2, 3). Using (6) in
(8) and (9) results in (12) and (13) given below.
𝐶𝑝1𝑑𝑢1
𝑑𝑡=
1
2𝑠11𝑔𝑉𝑚 tanh 𝛽 𝑎11𝑉1 + 𝑎12𝑉2– 𝑏1 + 1 +
1
2𝑠21 𝑔𝑉𝑚 tanh 𝛽 𝑎21𝑉1 + 𝑎22𝑉2– 𝑏2 + 1 +
𝑉1
𝑅11+
𝑉2
𝑅21 − 𝑢1
1
𝑅1 (12)
𝐶𝑝2𝑑𝑢2
𝑑𝑡=
1
2𝑠12𝑔𝑉𝑚 tanh 𝛽 𝑎11𝑉1 + 𝑎12𝑉2– 𝑏1 + 1 +
1
2𝑠22 𝑔𝑉𝑚 tanh 𝛽 𝑎21𝑉1 + 𝑎22𝑉2– 𝑏2 + 1 +
𝑉1
𝑅12+
𝑉2
𝑅22 − 𝑢2
1
𝑅2 (13)
CERA09
Moreover, it can be shown that the network in fig. 1 can be
associated with an Energy Function ‘E’ of the form
From (14), it follows that
𝜕𝐸
𝜕𝑉1= 𝑐11𝑉1 + 𝑐21𝑉2 +
1
2𝑉𝑚 𝑎11tanh 𝛽(𝑎11𝑉1 + 𝑎12𝑉2 − 𝑏1) +
1
2𝑉𝑚𝑎21tanh 𝛽(𝑎21𝑉1 + 𝑎22𝑉2 − 𝑏2) − 𝑢1
1
𝑅1 (15)
Also, if ‘E’ is the Energy Function, it must satisfy the
following condition [8].
𝜕𝐸
𝜕𝑉1= 𝐾 𝐶1
𝑑𝑢1
𝑑𝑡 (16)
Where ‘K’ is a constant of proportionality and has the
dimensions of resistance. Comparing (12) and (15) according
to (16) yields
𝑎11 = 𝑔𝐾𝑠11; 𝑎21 = 𝑔𝐾𝑠21 (17)
𝑐11 = 𝐾
𝑅11; 𝑐21 =
𝐾
𝑅21 (18)
The values of the current scaling factors (sji) of Fig. 1 can
be easily calculated by setting the value of ‘K’ to be equal to 1
𝑔 and then using (17). The values of the weight resistances
connected to the first neuron, R11 and R21, can be obtained by
using (18). Analysis on similar lines can be performed to
obtain the values of the synaptic weights for the second
neuron.
III. SIMULATION RESULTS
This section deals with the application of the proposed
network to task of minimizing the objective function
3𝑉12 + 4𝑉1𝑉2 + 5𝑉2
2 (19)
subject to
𝑉1 − 𝑉2 ≤ −1
𝑉1 + 𝑉2 ≤ 1
The values of resistances acting as the weights on the neurons
are obtained from (18). Towards that end, first the value of 𝑔
was computed for the DVCC which was found to be 1.091
milli-mhos. The corresponding values of the weight
resistances are found to be: R11 =305.53Ω; R21 =229.14Ω; R12
=229.14Ω; R22 =183.318Ω. The values of the current scaling
factors were found out to be: s11 = 1; s12 = -1; s21 = 1 ; s22 = 1.
The circuit for DVCC was taken from [7] and standard 0.5
micron CMOS parameters were used for simulation purposes.
For the opamp, use was made of the LMC7101A CMOS
opamp from National Semiconductor. The sub-circuit file for
this opamp is available in Orcad Model Library.
Routine mathematical analysis of (19, 20) yields: 𝑉1= -0.584,
𝑉2= 0.416. The results of PSPICE simulation are presented in
fig. 3. From the plots of the neuron output voltages, it can be
seen that V(1) = -0.58 V and V(2) = 0.41 V which are very
near to the algebraic solution thereby confirming the validity
of the approach.
Fig. 3 Simulation results for the chosen 2 – variable problem
IV. CONCLUSIONS
In this paper we have described a CMOS compatible
approach to solve a quadratic programming problem in 2
variables subject to 2 linear constraints, which uses 2-neurons
and 2-synapses. Each neuron requires one opamp and each
synapse is implemented using a dual output DVCC. This
results in significant reduction in hardware over the existing
schemes [2-6]. The proposed network was tested on a sample
problem of minimizing a quadratic function in 2 variables
and the simulation results confirm the validity of the
approach.
REFERENCES
[1] D.W. Tank, J.J. Hopfield, (1986) Simple neural optimization networks:
an A/D converter, signal decision network, and linear programming
circuit, IEEE Trans. Circ. Syst. CAS, 33, 533–541. [2] M.P. Kennedy, L.O. Chua, (1988) Neural networks for nonlinear
programming, IEEE Trans. Circ. Syst., 35, 554–562.
[3] C.Y. Maa, M.A. Shanblatt, (1992) A two-phase optimization neural network, IEEE Trans. Neural Network 3 (6), 580 – 594.
[4] X.Y. Wu, Y.S. Xia, J. Li, W.K. Chen, (1996) A high performance
neural network for solving linear and quadratic programming problems, IEEE Trans. Neural Networks, 7 (3), 643–651.
[5] Y. Xia, (1996) A new neural network for solving linear and quadratic
programming problems, IEEE Trans. Neural Networks 7 (6), 1544–1547.
[6] A. Malek, M. Alipour, (2007), Numerical solution for linear and
quadratic programming problems using a recurrent nueral network, Appl. Math. Computation, 192, 27 – 39.
[7] S. Maheshwari, (2008) A canonical voltage-controlled VM-APS with a
grounded capacitor,” Cir. Syst. Sig. Process, 27, 123-132. [8] S. A. Rahman, Jayadeva, S. C. Dutta Roy, (1999) Neural network
approach to graph colouring, Electronics Letters, 35, 1173 – 1175.
(20)
978-1-4244-8542-0/10/$26.00 ©2010 IEEE
A DVCC-based Non-Linear Analog Circuit for
Solving Linear Programming Problems
Mohd. Samar Ansari and Syed Atiqur Rahman
Abstract—This paper presents a neural circuit for solving
linear programming problem (LPP). The objective is to minimize
a first order cost function subject to linear constraints. The
dynamic analog circuit, consisting of N identical units for N
variable problem, can solve the general LPP and always
converges to the optimal solution in constant time, irrespective of
the initial conditions, which is of the order of its time constant.
The proposed circuit employs non-linear feedback, in the form of
Differential Voltage Current Conveyor (DVCC) based unipolar
comparators, to introduce transcendental terms in the energy
function ensuring fast convergence to the solution. Further, the
use of resistors to generate weighted inputs to the neurons is
avoided. Instead, DVCCs are utilized to directly generate the
required scaled currents. PSPICE simulation results are
presented for a chosen optimization problem and are found to
agree with the algebraic solution.
Index Terms— DVCC, Linear Programming, Neural network
applications, Nonlinear circuits, Optimization.
I. INTRODUCTION
inear programming is a considerable field of optimization
for several reasons. Many practical problems in
operations research can be expressed as linear programming
problems. Certain special cases of linear programming, such
as network flow problems and multi-commodity flow
problems are considered important enough to have generated
much research on specialized algorithms for their solution [1-
6]. A number of algorithms for other types of optimization
problems work by solving LP problems as sub-problems. In a
LPP problem, a linear function is optimized subject to certain
linear constraints. The general LPP problem can be stated as:
Minimize (1)
Subject to (2)
where Rn is a column vector of decision variables, R
n is
a column vector of cost coefficient, Rm is a column vector
of constant bounds, Rm x n
is a constraint coefficient matrix,
m n, and the superscript T denotes the transpose operator.
The m constraints specified by (2) define the feasible region in
the n dimensional Euclidean space in which the cost function
is required to be minimized [2].
Traditional methods for solving linear programming
problems typically involve an iterative process, but long
computational time limits their usage. There is an alternative
approach to solution of this problem.
Authors are with the Department of Electronics Engineering, Aligarh Muslim,
Aligarh, India (e-mail: [email protected]).
It is to exploit the artificial neural networks (ANN's) which
can be considered as an analog computer relying on a highly
simplified model of neurons [1]. ANN's have been applied to
several classes of constrained optimization problems and have
shown promise for solving such problems more effectively.
For example, the Hopfield neural network has proven to be a
powerful tool for solving some of the optimization problems.
Tank and Hopfield first proposed a neural network for
solving mathematical programming problems, where a
linear programming problem was mapped into a closed-loop
network [2]. Over the past two decades several neural-network
architectures for solving linear programming problems have
been proposed by Wang [3], Xia [4], Xia, Wang & Hung [5]
and Malek and Yari [6]. Wang‟s realization of a neural
network for solving linear programming problems utilizes 3n
amplifiers, n2+3n resistances and n capacitors. The circuit of
[4] requires 2m2+4mn amplifiers, 2m
2+4mn+3m+3 summers,
n+m integrators, and n simple limiters. The circuit of [5]
consists of m2+2mn amplifiers, 3n summers, n integrators, and
n simple limiters. And although no circuit implementation of
the proposed scheme is given in [6], it is evident that it would
not be simple to actually realize in hardware. Therefore, the
search still continues for a working circuit, to solve LPP,
which is feasible to implement in integrated forms and is fast
enough for utilization in real-time applications where the time-
to-solve is needed to be of the order of tenths of milliseconds
[5].
In this paper, a hardware solution to the problem of solving
a linear programming problem subject to linear inequality
constraints is presented. The proposed circuit employs only n
amplifiers, m DVCC-based unipolar comparators and
mn+m+n resistances. It is evident that the circuit complexity
of the proposed scheme is much reduced as compared to
existing ones [3-6].
The paper is organized as follows. Section II describes the
proposed neural network to minimize a first order polynomial
in 2 variables subject to 2 linear constraints. Section III
contains results of PSPICE simulation. Some concluding
remarks appear in section IV.
II. PROPOSED NETWORK
Let the function to be minimized be
(3)
Subject to
(4)
(5)
L
978-1-4244-8542-0/10/$26.00 ©2010 IEEE
Where V1, and V2 are the variables and aij, bi and cij are
constants. The proposed neural-network based circuit to
minimize (3) subject to (4, 5) is presented in Fig. 1.
b1
s11Ix
a11V1+ a12V2
s12Ix
s21Ix
a21V1+ a22V2
s22Ix
+
_
V1
+_
Rp2C2V2
Rp1 C1
R1
R2
A
B
P1
P2
x1 Z+
1
X
Y2
Y1
Z+
2
x1 Z+
1
X
Y2
Y1
Z+
2
b2
DVCC-based
unipolar comparator
DVCC-based
unipolar comparator
Fig. 1 The proposed neural network for solving a linear programming problem in 2 variables subject to 2 linear constraints
As can be seen from Fig. 1, individual equations from the set
of constraints are passed through non-linear synapses which
are realized using DVCC-based unipolar comparators.
x1
Z+
X
Y2
Y1
DVCC
Z-
VY1
VY2
VX
IX
IZ+
IZ-
Fig. 2 Symbolic Representation of DVCC
The symbolic diagram of the DVCC is shown in Fig. 2. The
device can be characterized by the following port relations:
[
]
[
]
[
]
(6)
Although both Z+ and Z- types of current outputs are
mentioned in (4), the DVCCs used in the proposed network
use only Z+ type of outputs. For the DVCC, unipolar
comparator action can be achieved by putting i.e. by
directly grounding the X-terminal [7] and operating the
DVCC from a unipolar biasing supply. For such a case, the
current in the X-port will saturate and can be written as
[ ( – ) ] (7)
Where β is the open-loop gain of the comparator (practically
very high) and are the saturated output current levels of
the comparator. Equation (7) can also be written in an
equivalent notation as
[ ( – ) ] (8)
where is the transconductance from the input voltage ports
(Y1 and Y2) to the output current port (X) and are the
biasing voltages of the DVCC-based unipolar comparator. By
virtue of DVCC action, this current will be transferred to the
Z+ ports as
[ ( – ) ] (9)
Where is the current scaling factor introduced during the
current conveying process from the X-port to the Z+ port.
Ideally, . However, by altering the aspect ratios of the
transistors used in the output stages of the DVCC, can be
made a scaled replica of . In that case, the value of will be
different from unity. This in property has been employed in
the design of the multi-output DVCCs needed for the
proposed circuit.
The outputs of the unipolar comparators are fed to neurons
having weighted inputs. These weighted neurons are realized
by using opamps where the scaled currents coming from
various unipolar comparators act as weights. Weighing is also
provided by the self-feedback resistances (R11 & R22 for
neurons 1 & 2 respectively) and the cross-feedback resistances
(R21 and R12 for neurons 1 & 2 respectively). Rpi and Cpi are
the input resistance and capacitance of the opamp
corresponding to the i-th neuron. These parasitic components
are included to model the dynamic nature of the opamp.
Node equation for nodes „A‟ and „B‟ give the equations of
motions of the first and second neurons as
[
] [
] (10)
where,
(11)
And,
[
] [
] (12)
where,
(13)
In equations (10) and (12) ui is the internal state of the i-th
neuron and is the current scaling factor at the j-th output of
i-th DVCC-based unipolar comparator. As is shown later in
this section, these weights are governed by the constraint
inequalities (4, 5). Using (8) in (10) and (12) results in (14)
and (15) given below.
[ ( – ) ]
[ ( – ) ] [
]
[
] (14)
[ ( – ) ]
[ ( – ) ] [
]
[
] (15)
978-1-4244-8542-0/10/$26.00 ©2010 IEEE
Moreover, it can be shown that the network of Fig. 1 may be
associated with an „energy‟ function of the form
( )
( – )
( – ) (16)
From (16), it follows that
( )
( ) (17)
Also, if „E‟ is the Energy Function, it must satisfy the
following condition [8].
(18)
Where „K‟ is a constant of proportionality and has the
dimensions of resistance. Substituting for
in (18) from
(14) and comparing with (17) yields
⁄
⁄ (19)
(20)
The values of the current scaling factors (sji) of Fig. 1 can
be easily calculated by choosing a suitable value of „K‟ and
then using (19). Analysis on similar lines can be performed to
obtain the values of the synaptic weights for the second
neuron and are presented in (21) and (22) below.
⁄
⁄ (21)
(22)
III. SIMULATION RESULTS
This section deals with the application of the proposed
network to task of minimizing the objective function
(23) subject to
The values of resistances acting as the weights on the neurons
are obtained from (20, 22). Towards that end, first the value of
was computed for the DVCC which was found to be 1.091
milli-mhos. Choosing ⁄ allows us to fix the current
scaling factors as given below:
The circuit for DVCC was taken from [7] and is reproduced in
Fig. 3 for reference. Standard 0.5 micron CMOS parameters
were used for simulation purposes. DVCC-based unipolar
comparators were obtained by operating the DVCCs using
+5V and 0.2V as VDD and VSS supply voltages respectively
instead of the ±2.5V supplies used in [7]. The bias voltage
VBB was kept at 0.05V.
Y1Y2
VBB
M1 M4M2 M3
M5 M6
M9 M10
M7
M11
M8
M12
XZ+
VSS
VDD
Fig. 3 CMOS implementation of the DVCC [7]
The DC transfer characteristics obtained are presented in Fig.
4 from which it may be observed that unipolar comparator
action is indeed realized. For the opamp, use was made of the
LMC7101A CMOS opamp from National Semiconductor.
The sub-circuit file for this opamp is available in Orcad Model
Library.
Fig. 4 DC transfer characteristics of the DVCC-based
unipolar comparator
Routine mathematical analysis of (23, 24) yields: = 0, =
1. The results of PSPICE simulation are presented in Fig. 5.
From the plots of the neuron output voltages, it can be seen
that V1 = - 9.97mV and V2 = 1.01 V which correspond closely
to the algebraic solution thereby confirming the validity of the
approach. The slight error in the obtained solutions may be
attributed to the non-ideal characteristics of the unipolar
comparators implemented using DVCCs, as depicted in Fig. 4.
(24)
978-1-4244-8542-0/10/$26.00 ©2010 IEEE
Fig. 5 Simulation results for the proposed circuit
IV. CONCLUSION
In this paper, a CMOS compatible approach to solve linear
programming problem in 2 variables subject to 2 linear
constraints employing 2-neurons and 2-synapses, is described.
Each neuron requires one opamp and each synapse is
implemented using a differential transconductance element
realized using DVCCs. This results in significant reduction in
hardware over the existing schemes [3-6]. The proposed
network was tested on a sample problem of minimizing a
linear function in 2 variables and the simulation results
confirm the validity of the approach.
V. REFERENCES
[1] J.J. Hopfield, D.W. Tank, “„Neural‟ computation of decisions
optimization problems,” Biological Cybern., 52, 141-152, 1985
[2] D. W. Tank and J. J. Hopfield, “Simple Neural Optimization Networks: An A/D Converter, Signal Decision Circuit, and A Linear Programming
Circuit,” IEEE Trans. Circuits and Systems, Vol. CAS-33, No.5, pp. 533-541, May 1986.
[3] J. Wang, “Analysis and Design of a Recurrent Neural Network for
Linear Programming”, IEEE Trans. Cir and Sys–1: Fundamental Theory and Applications, Vol.40, No.9, pp. 613-618, Sept.1993.
[4] Y. Xia, “A New Neural Network for Solving Linear and Quadratic
Programming Problems,” IEEE Trans. Neural Net., Vol.7, No.6, pp. 1544-1547, Nov 1996.
[5] Y. Xia, J. Wang and D. L. Hung, “Recurrent Neural Networks for
Solving Linear Inequalities and Equations,” IEEE Trans. Circuits and Systems–1: Fundamental Theory and Applications, Vol.46, No.4, pp.
452-462, April 1999.
[6] A. Malek and A. Yari., “Primal–dual solution for the linear programming problems using neural networks,” Applied Mathematics
and Computation, 167, 198–211, 2005.
[7] S. Maheshwari, (2008) A canonical voltage-controlled VM-APS with a grounded capacitor,” Cir. Syst. Sig. Process, 27, 123-132.
[8] S. A. Rahman, Jayadeva, S. C. Dutta Roy, “Neural network approach to
graph colouring,” Electronics Letters, 35, 1173 – 1175, 1999.