m15 part2 tier1 nonlinear - polytechnique montréal · 2006-10-23 · multivariable unconstrained...
TRANSCRIPT
Process Optimization
Click to edit Master title styleTier I: Mathematical Methods of
Optimization
Section 3:Nonlinear Programming
Click to edit Master title styleIntroduction to Nonlinear
Programming• We already talked about a basic aspect of
nonlinear programming (NLP) in the Introduction Chapter when we considered unconstrained optimization.
Click to edit Master title styleIntroduction to Nonlinear
Programming
• We optimized one-variable nonlinear functions using the 1st and 2nd derivatives.
• We will use the same concept here extended to functions with more than one variable.
Click to edit Master title styleMultivariable Unconstrained
Optimization
• For functions with one variable, we use the 1st
and 2nd derivatives.
• For functions with multiple variables, we use identical information that is the gradient and the Hessian.
• The gradient is the first derivative with respect to all variables whereas the Hessian is the equivalent of the second derivative
Click to edit Master title styleThe Gradient
• Review of the gradient (∇):For a function “f ”, of variables x1, x2, …, xn:
⎥⎦
⎤⎢⎣
⎡∂∂
∂∂
∂∂
=∇nx
fxf
xff K
21
Example: 231
321 )(3)(215 xxxxf −+=
[ ]312
22
3 6)(6)(315 xxxxf −−=∇
Click to edit Master title styleThe Hessian
• The Hessian (∇2) of f(x1, x2, …, xn) is:
⎥⎥⎥⎥⎥⎥⎥⎥
⎦
⎤
⎢⎢⎢⎢⎢⎢⎢⎢
⎣
⎡
∂∂
∂∂∂
∂∂∂
∂∂∂
∂∂
∂∂∂
∂∂∂
∂∂∂
∂∂
=∇
2
2
2
2
1
2
2
2
22
2
12
21
2
21
2
21
2
2
nnn
n
n
xf
xxf
xxf
xxf
xf
xxf
xxf
xxf
xf
f
L
MOMM
K
K
Click to edit Master title styleHessian Example
• Example (from previously):2
313
21 )(3)(215 xxxxf −+=
⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡
−−
−=∇
13
2
32
6060120600
xxx
xf
[ ]312
22
3 6)(6)(315 xxxxf −−=∇
Click to edit Master title styleUnconstrained Optimization
The optimization procedure for multivariable functions is:
1. Solve for the gradient of the function equal to zero to obtain candidate points.
2. Obtain the Hessian of the function and evaluate it at each of the candidate points
• If the result is “positive definite” (defined later) then the point is a local minimum.
• If the result is “negative definite” (defined later) then the point is a local maximum.
Click to edit Master title stylePositive/Negative Definite
• A matrix is “positive definite” if all of the eigenvalues of the matrix are positive (> 0)
• A matrix is “negative definite” if all of the eigenvalues of the matrix are negative (< 0)
Click to edit Master title stylePositive/Negative Semi-definite
• A matrix is “positive semi-definite” if all of the eigenvalues are non-negative (≥ 0)
• A matrix is “negative semi-definite” if all of the eigenvalues are non-positive (≤ 0)
Click to edit Master title styleExample Matrix
Given the matrix A:
⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡−−−=211175
542A
The eigenvalues of A are:22 −=λ702.31 −=λ 702.23 =λ
This matrix is negative definite
Click to edit Master title styleUnconstrained NLP Example
Consider the problem:Minimize f(x1, x2, x3) = (x1)2 + x1(1 – x2) + (x2)2
– x2x3 + (x3)2 + x3
First, we find the gradient with respect to xi:
1 2
1 2 3
2 3
2 122 1
x xf x x x
x x
+ −⎡ ⎤⎢ ⎥∇ = − + −⎢ ⎥⎢ ⎥− + +⎣ ⎦
Click to edit Master title styleUnconstrained NLP Example
Next, we set the gradient equal to zero:
0=∇f⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡=
⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡
++−−+−−+
000
12212
32
321
21
xxxx
xx⇒
So, we have a system of 3 equations and 3 unknowns. When we solve, we get:
⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡
−−−
=⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡=
111
3
2
1
xxx
x
Click to edit Master title styleUnconstrained NLP Example
So we have only one candidate point to check.
Find the Hessian:
⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡
−−−
−=∇
210121
0122 f
Click to edit Master title styleUnconstrained NLP Example
The eigenvalues of this matrix are:
414.31 =λ 586.02 =λ 23 =λ
All of the eigenvalues are > 0, so the Hessian is positive definite.
So, the point is a minimum⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡
−−−
=⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡=
111
3
2
1
xxx
x
Click to edit Master title styleUnconstrained NLP Example
Unlike in Linear Programming, unless we know the shape of the function being minimized or can determine whether it is convex, we cannot tell whether this point is the global minimum or if there are function values smaller than it.
Click to edit Master title styleMethod of Solution
• In the previous example, when we set the gradient equal to zero, we had a system of 3 linear equations & 3 unknowns.
• For other problems, these equations could be nonlinear.
• Thus, the problem can become trying to solve a system of nonlinear equations, which can be very difficult.
Click to edit Master title styleMethod of Solution
• To avoid this difficulty, NLP problems are usually solved numerically.
• We will now look at examples of numerical methods used to find the optimum point for single-variable NLP problems. These and other methods may be found in any numerical methods reference.
Click to edit Master title styleNewton’s Method
When solving the equation f ′(x) = 0 to find a minimum or maximum, one can use the iteration step:
)()(
''
'1
k
kkk
xfxfxx −=+
where k is the current iteration. Iteration is continued until |xk+1 – xk| < ε where ε is some specified tolerance.
Click to edit Master title styleNewton’s Method Diagram
Newton’s Method approximates f′ (x) as a straight line at xk and obtains a new point (xk+1), which is used to approximate the function at the next iteration. This is carried on until the new point is sufficiently close to x*.
xx* xk+1 xk
Tangent of f′ (x) at xk
f′ (x)
Click to edit Master title styleNewton’s Method Comments
• One must ensure that f (xk+1) < f (xk) for finding a minimum and f (xk+1) > f (xk) for finding a maximum.
• Disadvantages:– Both the first and second derivatives must be
calculated– The initial guess is very important – if it is not
close enough to the solution, the method may not converge
Click to edit Master title styleRegula-Falsi Method
This method requires two points, xa & xb that bracket the solution to the equation f ′(x) = 0.
)()()()(
''
'
ab
abbbc
xfxfxxxfxx
−−⋅
−=
where xc will be between xa & xb. The next interval will be xc and either xa or xb, whichever has the sign opposite of xc.
Click to edit Master title styleRegula-Falsi Diagram
The Regula-Falsi method approximates the function f′ (x) as a straight line and interpolates to find the root.
xxa
xb
xc
x*
f′ (x)
Click to edit Master title styleRegula-Falsi Comments
• This method requires initial knowledge of two points bounding the solution
• However, it does not require the calculation of the second derivative
• The Regula-Falsi Method requires slightly more iterations to converge than the Newton’s Method
Click to edit Master title styleMultivariable Optimization
• Now we will consider unconstrained multivariable optimization
• Nearly all multivariable optimization methods do the following:
1. Choose a search direction dk
2. Minimize along that direction to find a new point:
where k is the current iteration number and αk
is a positive scalar called the step size.
kkkk dxx α+=+1
Click to edit Master title styleThe Step Size
• The step size, αk, is calculated in the following way:
• We want to minimize the function f(xk+1) = f(xk +αkdk) where the only variable is αk
because xk & dk are known.• We set and solve for αk
using a single-variable solution method such as the ones shown previously.
( ) 0d
d=
+k
kkkfαα dx
Click to edit Master title styleSteepest Descent Method
• This method is very simple – it uses the gradient (for maximization) or the negative gradient (for minimization) as the search direction:
)( kk f xd ∇⎭⎬⎫
⎩⎨⎧−+
= for⎭⎬⎫
⎩⎨⎧
minmax
So, )(1 kkkk f xxx ∇⎭⎬⎫
⎩⎨⎧−+
=+ α
Click to edit Master title styleSteepest Descent Method
• Because the gradient is the rate of change of the function at that point, using the gradient (or negative gradient) as the search direction helps reduce the number of iterations needed
x1
x2 f(x) = 5f(x) = 20
f(x) = 25
xk
∇f(xk)
-∇f(xk)
Click to edit Master title styleSteepest Descent Method Steps
So the steps of the Steepest Descent Method are:
1. Choose an initial point x0
2. Calculate the gradient ∇f(xk) where k is the iteration number
3. Calculate the search vector:4. Calculate the next x:
Use a single-variable optimization method to determine αk.
)( kk f xd ±∇=kkkk dxx α+=+1
Click to edit Master title styleSteepest Descent Method Steps
5. To determine convergence, either use some given tolerance ε1 and evaluate:
for convergenceOr, use another tolerance ε2 and evaluate:
for convergence
11 )()( ε<−+ kk ff xx
2)( ε<∇ kf x
Click to edit Master title styleConvergence
• These two criteria can be used for any of the multivariable optimization methods discussed here
Recall: The norm of a vector, ||x|| is given by:22
22
1 )()()( nT xxx +++=⋅= Lxxx
Click to edit Master title styleSteepest Descent Example
Let’s solve the earlier problem with the Steepest Descent Method:
Minimize f(x1, x2, x3) = (x1)2 + x1(1 – x2) + (x2)2
– x2x3 + (x3)2 + x3
Let’s pick
⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡=
000
0x
Click to edit Master title styleSteepest Descent Example
[ ]10000001)0(2)( 00 ++−−+−−+−=−∇= xd f
[ ]122)1(2)( 3232121 ++−−+−−+=∇ xxxxxxxf x
[ ] [ ]101101 −−=−=
[ ] [ ]101000 01 −−⋅+= αx
Now, we need to determine α0
Click to edit Master title styleSteepest Descent Example
2)(4)( 00
1
−= ααd
df x
)()(00)1)(()()( 0200201 αααα −++−+−+=xf
)(2)(2 020 αα −=
Now, set equal to zero and solve:
21
420 ==α2)(4 0 =α ⇒
Click to edit Master title styleSteepest Descent Example
So,[ ] [ ]101000 01 −−⋅+= αx
[ ] ⎥⎦⎤
⎢⎣⎡ −−+=
210
21000
⎥⎦⎤
⎢⎣⎡ −−=∴
210
211x
Click to edit Master title styleSteepest Descent Example
⎥⎦⎤
⎢⎣⎡ +−++++−−=−∇= 110
210
21011)( 11 xd f
[ ]0101 −=∴d
Take the negative gradient to find the next search direction:
Click to edit Master title styleSteepest Descent Example
[ ]010210
21 12 −⋅+⎥⎦
⎤⎢⎣⎡ −−= αx
⎥⎦⎤
⎢⎣⎡ −−−=
21
21 1α
Update the iteration formula:
Click to edit Master title styleSteepest Descent Example
1)(2)( 11
1
−= ααd
df x
( )21
41
21)()(1
21
41)( 12112 −+⎟
⎠⎞
⎜⎝⎛−++⎟
⎠⎞
⎜⎝⎛−+= αααxf
21)( 121 −−= αα
Insert into the original function & take the derivative so that we can find α1:
Click to edit Master title styleSteepest Descent Example
Now we can set the derivative equal to zero and solve for α1:
211 =α1)(2 1 =α ⇒
Click to edit Master title styleSteepest Descent Example
Now, calculate x2:
[ ]010210
21 12 −⋅+⎥⎦
⎤⎢⎣⎡ −−= αx
⎥⎦⎤
⎢⎣⎡ −+⎥⎦
⎤⎢⎣⎡ −−= 0
210
210
21
⎥⎦⎤
⎢⎣⎡ −−−=∴
21
21
212x
Click to edit Master title styleSteepest Descent Example
So,
⎥⎦⎤
⎢⎣⎡ +−+−++−−=−∇= 11
21
211
21
2111)( 22 xd f
⎥⎦⎤
⎢⎣⎡ −−=∴
210
212d
⎥⎦⎤
⎢⎣⎡ −−⋅+⎥⎦
⎤⎢⎣⎡ −−−=
210
21
21
21
21 23 αx
⎥⎦⎤
⎢⎣⎡ +−−+−= )1(
21
21)1(
21 22 αα
Click to edit Master title styleSteepest Descent Example
23)1()( 2
2
3
−+= ααd
df x
41)1(
23)1(
21)( 2223 ++−+= ααxf
Set the derivative equal to zero and solve:
212 =α
23)1( 2 =+α ⇒
Find α2:
Click to edit Master title styleSteepest Descent Example
Calculate x3:
⎥⎦⎤
⎢⎣⎡ −−⋅+⎥⎦
⎤⎢⎣⎡ −−−=
210
21
21
21
21 23 αx
⎥⎦⎤
⎢⎣⎡ −−+⎥⎦
⎤⎢⎣⎡ −−−=
410
41
21
21
21
⎥⎦⎤
⎢⎣⎡ −−−=∴
43
21
433x
Click to edit Master title styleSteepest Descent Example
⎥⎦⎤
⎢⎣⎡ −=⎥⎦
⎤⎢⎣⎡−=−∇= 0
2100
210)( 33 xd f
⎥⎦⎤
⎢⎣⎡ −⋅+⎥⎦
⎤⎢⎣⎡ −−−= 0
210
43
21
43 34 αx
⎥⎦⎤
⎢⎣⎡ −+−−=
43)1(
21
43 3α
Find the next search direction:
Click to edit Master title styleSteepest Descent Example
089)1(
21)( 3
3
4
=−+= ααd
df x
23)(
23)1(
41)( 3234 −−+= ααxfFind α3:
453 =⇒α
Click to edit Master title styleSteepest Descent Example
So, x4 becomes:
⎥⎦⎤
⎢⎣⎡ −+⎥⎦
⎤⎢⎣⎡ −−−= 0
850
43
21
434x
⎥⎦⎤
⎢⎣⎡ −−−=∴
43
89
434x
Click to edit Master title styleSteepest Descent Example
⎥⎦⎤
⎢⎣⎡ −−=⎥⎦
⎤⎢⎣⎡ −−=−∇=
85
43
85
85
43
85)( 44 xd f
⎥⎦⎤
⎢⎣⎡ −−⋅+⎥⎦
⎤⎢⎣⎡ −−−=
85
43
85
43
89
43 45 αx
⎥⎦⎤
⎢⎣⎡ +−−−+−= )
253(
41)
23(
43)
253(
41 444 ααα
The next search direction:
Click to edit Master title styleSteepest Descent Example
03243
1673)( 4
4
5
=−= ααd
df x
6451
3243)(
3273)( 4245 −−= ααxf
146434 =α⇒
Find α4:
Click to edit Master title styleSteepest Descent Example
Update x5:
⎥⎦⎤
⎢⎣⎡ −−⋅+⎥⎦
⎤⎢⎣⎡ −−−=
85
43
85
14643
43
89
435x
⎥⎦⎤
⎢⎣⎡ −−−=∴
11681091
7366
116810915x
Click to edit Master title styleSteepest Descent Example
Let’s check to see if the convergence criteria is satisfied
Evaluate ||∇f(x5)||:
⎥⎦⎤
⎢⎣⎡=∇
58421
58435
58421)( 5xf
( ) ( ) ( ) 0786.058421
58435
58421)(
2225 =++=∇ xf
Click to edit Master title styleSteepest Descent Example
So, ||∇f(x5)|| = 0.0786, which is very small and we can take it to be close enough to zero for our example
Notice that the answer of
⎥⎦⎤
⎢⎣⎡ −−−=
11681091
7366
11681091x
is very close to the value of that we obtained analytically
[ ]111* −−−=x
Click to edit Master title styleQuadratic Functions
• Quadratic functions are important for the next method we will look at
• A quadratic function can be written in the form: xTQx where x is the vector of variables and Q is a matrix of coefficients
Example:[ ] )2(x x2x )(x2
210121
0122
2212
1
3
2
1
321 +−=⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡
⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡
−−−
−
xxx
xxx
)2(x x2x– 2332 +
Click to edit Master title styleConjugate Gradient Method
• The Conjugate Gradient Method has the property that if f(x) is quadratic, it will take exactly n iterations to converge, where n is the number of variables in the x vector
• Although it works especially well with quadratic functions, this method will work with non-quadratic functions also
Click to edit Master title styleConjugate Gradient Steps
1. Choose a starting point x0 and calculate f(x0). Let d0 = -∇f(x0)
2. Calculate x1 using:Find α0 by performing a single-variable optimization on f(x0 +α0d0) using the methods discussed earlier. (See illustration after algorithm explanation)
0001 dxx α+=
Click to edit Master title styleConjugate Gradient Steps
3. Calculate f(x1) and ∇f(x1). The new search direction is calculated using the equation:
)()()()()( 00
11011
xxxxdxd
fffff T
T
∇⋅∇∇⋅∇
+−∇=
This can be generalized for the kth iteration:
)()()()()(
1111
kkT
kkTkkk
fffff
xxxxdxd
∇⋅∇∇⋅∇
+−∇=++
++
Click to edit Master title styleConjugate Gradient Steps
4. Use either of the two methods discussed earlier to determine tolerance:
11 )()( ε<−+ kk ff xx
2)( ε<∇ kf x
Or,
Click to edit Master title styleNumber of Iterations
• For quadratic functions, this method will converge in n iterations (k = n)
• For non-quadratic functions, after niterations, the algorithm cycles again with dn+1 becoming d0.
Click to edit Master title styleStep Size for Quadratic Functions
• When optimizing the step size, we can approximate the function to be optimized in the following manner:
)()()()()()( 221 kkTkkkTkkk ffff dxddxxdx αααα ⋅∇⋅+⋅∇+≈+
• For a quadratic function, this is not an approximation – it is exact
Click to edit Master title styleStep Size for Quadratic Functions
We take the derivative of that function with respect to α and set it equal to zero:
0)()()()( 2 =⋅∇⋅+⋅∇=+ kkTkkkT
kk
ffd
df dxddxdx ααα
The solution to this equation is:
kkTk
kkToptk
ff
dxddx⋅∇⋅
⋅∇−=
)()()(
2,α
Click to edit Master title styleStep Size for Quadratic Functions
• So, for the problem of optimizing a quadratic function,
kkTk
kkToptk
ff
dxddx⋅∇⋅
⋅∇−=
)()()(
2,α
is the optimum step size.• For a non-quadratic function, this is an
approximation of the optimum step size.
Click to edit Master title styleMultivariable Newton’s Method
We can approximate the gradient of f at a point x0 by:
)()()()( 0020 xxxxx −⋅∇+∇≈∇ fff
We can set the right-hand side equal to zero and rearrange to give:
[ ] )()( 01020 xxxx ff ∇⋅∇−=−
Click to edit Master title styleMultivariable Newton’s Method
We can generalize this equation to give an iterative expression for the Newton’s Method:
[ ] )()( 121 kkkk ff xxxx ∇⋅∇−=−+
where k is the iteration number
Click to edit Master title styleNewton’s Method Steps
1. Choose a starting point, x0
2. Calculate ∇f(xk) and ∇2f(xk)3. Calculate the next x using the equation
4. Use either of the convergence criteria discussed earlier to determine convergence. If it hasn’t converged, return to step 2.
[ ] )()( 121 kkkk ff xxxx ∇⋅∇−=−+
Click to edit Master title styleComments on Newton’s Method
• We can see that unlike the previous two methods, Newton’s Method uses both the gradient and the Hessian
• This usually reduces the number of iterations needed, but increases the computation needed for each iteration
• So, for very complex functions, a simpler method is usually faster
Click to edit Master title styleNewton’s Method Example
For an example, we will use the same problem as before:
Minimize f(x1, x2, x3) = (x1)2 + x1(1 – x2) + (x2)2
– x2x3 + (x3)2 + x3
[ ]12212)( 3232121 ++−−+−+−=∇ xxxxxxxf x
Click to edit Master title styleNewton’s Method Example
The Hessian is:
⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡
−−−
−=∇
210121
012)(2 xf
And we will need the inverse of the Hessian:
[ ]⎥⎥⎥⎥
⎦
⎤
⎢⎢⎢⎢
⎣
⎡
=⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡
−−−
−=∇
−
−
43
21
41
2112
14
12
14
3
210121
012)(
1
12 xf
Click to edit Master title styleNewton’s Method Example
So, pick⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡=
000
0x
Calculate the gradient for the 1st iteration:
[ ]100000100)( 0 ++−−+−+−=∇ xf
[ ]101)( 0 =∇⇒ xf
Click to edit Master title styleNewton’s Method Example
So, the new x is:11 0 2 0 0( ) ( )
3 1 14 2 40 1
1 1 0 1 02 20 131 1
4 2 4
f f−
⎡ ⎤= − ∇ ⋅∇⎣ ⎦⎡ ⎤
⎡ ⎤ ⎡ ⎤⎢ ⎥⎢ ⎥ ⎢ ⎥⎢ ⎥= − ⋅⎢ ⎥ ⎢ ⎥⎢ ⎥⎢ ⎥ ⎢ ⎥⎢ ⎥⎣ ⎦ ⎣ ⎦
⎢ ⎥⎣ ⎦
x x x x
⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡
−−−
=∴111
1x
Click to edit Master title styleNewton’s Method Example
Now calculate the new gradient:
[ ] [ ]000121121112)( 1 =+−+−++−=∇ xf
Since the gradient is zero, the method has converged
Click to edit Master title styleComments on Example
• Because it uses the 2nd derivative, Newton’s Method models quadratic functions exactly and can find the optimum point in one iteration.
• If the function had been a higher order, the Hessian would not have been constant and it would have been much more work to calculate the Hessian and take the inverse for each iteration.
Click to edit Master title styleConstrained Nonlinear
Optimization• Previously in this chapter, we solved NLP
problems that only had objective functions, with no constraints.
• Now we will look at methods on how to solve problems that include constraints.
Click to edit Master title styleNLP with Equality Constraints
• First, we will look at problems that only contain equality constraints:
Minimize f(x) x = [x1 x2 … xn]Subject to: hi(x) = bi i = 1, 2, …, m
Click to edit Master title styleIllustration
Consider the problem:Minimize x1 + x2
Subject to: (x1)2 + (x2)2 – 1 = 0
The feasible region is a circle with a radius of one. The possible objective function curves are lines with a slope of -1. The minimum will be the point where the lowest line still touches the circle.
Click to edit Master title styleGraph of Illustration
f(x) = 1
f(x) = 0
f(x) = -1.414
Feasible region
⎥⎦
⎤⎢⎣
⎡=
707.0707.0*x
)(xf∇
The gradient of f points in the direction of
increasing f
Click to edit Master title styleMore on the Graph
• Since the objective function lines are straight parallel lines, the gradient of f is a straight line pointing toward the direction of increasing f, which is to the upper right
• The gradient of h will be pointing out from the circle and so its direction will depend on the point at which the gradient is evaluated.
Click to edit Master title styleFurther Details
f(x) = 1
f(x) = 0
f(x) = -1.414
Feasible region
⎥⎦
⎤⎢⎣
⎡=
707.0707.0*x
)( *xh∇
)( *xf∇
⎥⎦
⎤⎢⎣
⎡=
011x
)( 1xf∇
Tangent Plane
)( 1xh∇x1
x2
Click to edit Master title styleConclusions
• At the optimum point, ∇f(x) is perpendicular to ∇h(x)
• As we can see at point x1, ∇f(x) is not perpendicular to ∇h(x) and we can move (down) to improve the objective function
• We can say that at a max or min, ∇f(x)must be perpendicular to ∇h(x)– Otherwise, we could improve the objective
function by changing position
Click to edit Master title styleFirst Order Necessary Conditions
So, in order for a point to be a minimum (or maximum), it must satisfy the following equation:
This equation means that ∇f(x*) and ∇h(x*)must be in exactly opposite directions at a minimum or maximum point
0)()( *** =∇⋅+∇ xx hf λ
Click to edit Master title styleThe Lagrangian Function
To help in using this fact, we introduce the Lagrangian Function, L(x,λ):
)()(),( xxx hfL ⋅+= λλ
Review: The notation ∇x f(x,y) means the gradient of f with respect to x.
So, )()(),( xxx xxx hfL ∇⋅+∇=∇ λλ
Click to edit Master title styleFirst Order Necessary Conditions
• So, using the new notation to express the First Order Necessary Conditions (FONC), if x* is a minimum (or maximum) then
( ) 0),( ** ,=∇
λλ
xx xL
and to ensure feasibility.0)( * =xh
Click to edit Master title styleFirst Order Necessary Conditions
• Another way to think about it is that the one Lagrangian function includes all information about our problem
• So, we can treat the Lagrangian as an unconstrained optimization problem with variables x1, x2, …, xn and λ1, λ2, …, λm.
We can solve it by solving the equations
0x=
∂∂L
& 0λ=
∂∂L
Click to edit Master title styleUsing the FONC
Using the FONC for the previous example,
)()(),( xxx hfL ⋅+= λλ( ) ( )( )12
22
121 −+⋅++= xxxx λ
⎥⎦
⎤⎢⎣
⎡=
⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡
∂∂
∂∂
=∇00
),(2
1
xL
xL
Lx λx
And the first FONC equation is:
Click to edit Master title styleFONC Example
This becomes:
&
021 11
=+=∂∂ xxL λ
021 22
=+=∂∂ xxL λ
The feasibility equation is:
or,
( ) ( ) 0122
21 =−+ xx
( ) ( ) 0122
21 =−+=
∂∂ xxLλ
Click to edit Master title styleFONC Example
So, we have three equations and three unknowns.
When they are solved simultaneously, we obtain
707.021 ±== xx & 707.0m=λ
We can see from the graph that positive x1 & x2 corresponds to a maximum while negative x1 & x2 corresponds to the minimum.
Click to edit Master title styleFONC Observations
• If you go back to the LP Chapter and look at the mathematical definition of the KKT conditions, you may notice that they look just like our FONC that we just used
• This is because it is the same concept• We simply used a slightly different
derivation this time but obtained the same result
Click to edit Master title styleLimitations of FONC
• The FONC do not guarantee that the solution(s) will be minimums/maximums.
• As in the case of unconstrained optimization, they only provide us with candidate points that need to be verified by the second order conditions.
• Only if the problem is convex do the FONC guarantee the solutions will be extreme points.
Click to edit Master title styleSecond Order Necessary
Conditions (SONC)For where
and for y where
)()(),( 222 xxx hfLx ∇⋅+∇=∇ λλ),(2 λxLx∇
0y
x
xyxJ
x
=⋅
⎥⎥⎥⎥
⎦
⎤
⎢⎢⎢⎢
⎣
⎡
∂∂
∂∂
=⋅
)(
1
*
*
)(m
hh
h
M
If x* is a local minimum, then 0),( *2 ≥⋅∇⋅ yxy λLxT
Click to edit Master title styleSecond Order Sufficient Conditions
(SOSC)
• y can be thought of as being a tangent plane as in the graphical example shown previously– Jh is just the gradients of each h(x) equation
and we saw in the example that the tangent plane must be perpendicular to ∇h(x) and that is why
0),( *2 >⋅∇⋅ yxy λLxT
0yJ =⋅)(xh
Click to edit Master title styleThe y Vector
x1
x3
x2
Tangent Plane(all possible y vectors))(xh∇
0)( =xh
*x
The tangent plane is the location of all yvectors and intersects with x*
It must be orthogonal (perpendicular) to ∇h(x)
Click to edit Master title styleMaximization Problems
• The previous definitions of the SONC & SOSC were for minimization problems
• For maximization problems, the sense of the inequality sign will be reversed
For maximization problems:SONC:
SOSC:
0),( *2 ≤⋅∇⋅ yxy λLxT
0),( *2 <⋅∇⋅ yxy λLxT
Click to edit Master title styleNecessary & Sufficient
• The necessary conditions are required for a point to be an extremum but even if they are satisfied, they do not guarantee that the point is an extremum.
• If the sufficient conditions are true, then the point is guaranteed to be an extremum. But if they are not satisfied, this does not mean that the point is not an extremum.
Click to edit Master title styleProcedure
1. Solve the FONC to obtain candidate points.
2. Test the candidate points with the SONC– Eliminate any points that do not satisfy the
SONC3. Test the remaining points with the SOSC
– The points that satisfy them are min/max’s – For the points that do not satisfy, we cannot
say whether they are extreme points or not
Click to edit Master title styleProblems with Inequality
ConstraintsWe will consider problems such as:Minimize f(x)Subject to: hi(x) = 0 i = 1, …, m
& gj(x) ≤ 0 j = 1, …, pAn inequality constraint, gj(x) ≤ 0 is called “active” at x* if gj(x*) = 0.Let the set I(x*) contain all the indices of the active constraints at x*:
0)( * =xjg for all j in set I(x*)
Click to edit Master title styleLagrangian for Equality &
Inequality Constraint ProblemsThe Lagrangian is written:
We use λ’s for the equalities & μ’s for the inequalities.
∑∑==
⋅+⋅+=p
jjji
m
ii ghfL
11)()()(),,( xxxμλx μλ
Click to edit Master title styleFONC for Equality & Inequality
ConstraintsFor the general Lagrangian, the FONC
become
∑∑==
=∇⋅+∇⋅+∇=∇p
jjji
m
ii ghfL
1
*
1
* *)(*)(*)(*)*,*,( 0xxxμλx μλ
and the complementary slackness condition:
[ ] ,0)( ** =⋅ xjj gμ,0* ≥jμ pj K,1=
Click to edit Master title styleSONC for Equality & Inequality
ConstraintsThe SONC (for a minimization problem) are:
where as before.
0yxy x ≥⋅∇⋅ ),,( ***2 μλLT
0yxJ =⋅)( *
This time, J(x*) is the matrix of the gradients of all the equality constraints and only the inequality constraints that are active at x*.
Click to edit Master title styleSOSC for Equality & Inequality
Constraints• The SOSC for a minimization problem with
equality & inequality constraints are:
0yxy >⋅∇⋅ ),,( ***2 μλLxT
Click to edit Master title styleGeneralized Lagrangian Example
• Solve the problem:Minimize f(x) = (x1 – 1)2 + (x2)2
Subject to: h(x) = (x1)2 + (x2)2 + x1 + x2 = 0g(x) = x1 – (x2)2 ≤ 0
The Lagrangian for this problem is:
( ) ( ) ( ){ }2122
12
22
1 2)1(),,( xxxxxxL +++⋅++−= λμλx
( ){ }221 xx −⋅+ μ
Click to edit Master title styleGeneralized Lagrangian Example
• The first order necessary conditions:
( ) 0212 111
=++⋅⋅+−=∂∂ μλλ xxxL
0222 2222
=⋅⋅−+⋅⋅+=∂∂ xxxxL μλλ
( ) ( ) 0212
22
1 =+++=∂∂ xxxxLλ
( )( ) 0221 =−⋅ xxμ
Click to edit Master title styleGeneralized Lagrangian Example
Solving the 4 FONC equations, we get 2 solutions:
⎥⎦
⎤⎢⎣
⎡−
=4534.0
2056.0)1(x 9537.0=μ45.0=λ &
and ⎥⎦
⎤⎢⎣
⎡=
00)2(x 2−=μ0=λ &2)
1)
Click to edit Master title styleGeneralized Lagrangian Example
Now try the SONC at the 1st solution:Both h(x) & g(x) are active at this point (they both equal zero). So, the Jacobian is the gradient of both functions evaluated at x(1):
( )( ) ⎥⎦
⎤⎢⎣
⎡−−
=⎥⎦
⎤⎢⎣
⎡−
++=
9068.010932.0411.1
211212
)1(2
211
x
xJx
xx
Click to edit Master title styleGeneralized Lagrangian Example
The only solution to the equation:0yxJ =⋅)( )1(
is: ⎥⎦
⎤⎢⎣
⎡=
0000
y
And the Hessian of the Lagrangian is:
⎥⎦
⎤⎢⎣
⎡=⎥
⎦
⎤⎢⎣
⎡−+
+=∇
993.0009.2
2220022
)1(
2
x
x μλλ
L
Click to edit Master title styleGeneralized Lagrangian Example
So, the SONC equation is:
0≥⎥⎦
⎤⎢⎣
⎡=⎥
⎦
⎤⎢⎣
⎡⋅⎥⎦
⎤⎢⎣
⎡⋅⎥⎦
⎤⎢⎣
⎡0000
0000
993.0009.2
0000
This inequality is true, so the SONC is satisfied for x(1) and it is still a candidate point.
Click to edit Master title styleGeneralized Lagrangian Example
The SOSC equation is:
And we just calculated the left-hand side of the equation to be the zero matrix. So, in our case for x2:
0yxy x >⋅∇⋅ ),,( ***2 μλLT
0yyxx >⎥
⎦
⎤⎢⎣
⎡=⋅∇⋅
0000
)1(
2LT
So, the SOSC are not satisfied.
Click to edit Master title styleGeneralized Lagrangian Example
For the second solution:Again, both h(x) & g(x) are active at this point. So, the Jacobian is:
( ) ⎥⎦
⎤⎢⎣
⎡−
=⎥⎦
⎤⎢⎣
⎡−
++=
0111
211212
)2(2
21)2(
x
xJx
xx
Click to edit Master title styleGeneralized Lagrangian Example
The only solution to the equation:0yxJ =⋅)( )2(
is: ⎥⎦
⎤⎢⎣
⎡=
0000
y
And the Hessian of the Lagrangian is:
⎥⎦
⎤⎢⎣
⎡−
=⎥⎦
⎤⎢⎣
⎡−+
+=∇
2002
2220022
)2(
2
x
x μλλ
L
Click to edit Master title styleGeneralized Lagrangian Example
So, the SONC equation is:
0≥⎥⎦
⎤⎢⎣
⎡=⎥
⎦
⎤⎢⎣
⎡⋅⎥⎦
⎤⎢⎣
⎡−
⋅⎥⎦
⎤⎢⎣
⎡0000
0000
2002
0000
This inequality is true, so the SONC is satisfied for x(2) and it is still a candidate point
Click to edit Master title styleGeneralized Lagrangian Example
The SOSC equation is:
And we just calculated the left-hand side of the equation to be the zero matrix. So, in our case for x2:
0yxy x >⋅∇⋅ ),,( ***2 μλLT
0yyxx >⎥
⎦
⎤⎢⎣
⎡=⋅∇⋅
0000
)2(
2LT
So, the SOSC are not satisfied.
Click to edit Master title styleExample Conclusions
• So, we can say that both x(1) & x(2) may be local minimums, but we cannot be sure because the SOSC are not satisfied for either point.
Click to edit Master title styleNumerical Methods
• As you can see from this example, the most difficult step is to solve a system of nonlinear equations to obtain the candidate points.
• Instead of taking gradients of functions, automated NLP solvers use various methods to change a general NLP into an easier optimization problem.
Click to edit Master title styleExcel Example
Let’s solve the previous example with Excel:
Minimize f(x) = (x1 – 1)2 + (x2)2
Subject to: h(x) = (x1)2 + (x2)2 + x1 + x2 = 0g(x) = x1 – (x2)2 ≤ 0
Click to edit Master title styleExcel Example
We enter the objective function and constraint equations into the spreadsheet:
Variables:x1 x20 0
Value LimitObjective Function: =(A3-1)^2+B3^2
Constraint 1: =A3^2+B3^2+A3+B3 0Constraint 2: =A3-B3^2 0
Click to edit Master title styleExcel Example
Now, open the solver dialog box under the Tools menu and specify the objective function value as the target cell and choose the Min option. As it is written, A3 & B3 are the variable cells. And the constraints should be added – the equality constraint and the ≤ constraint.
Click to edit Master title styleExcel Example
The solver box should look like the following:
Click to edit Master title styleExcel Example
• This is a nonlinear model, so unlike the examples in the last chapter, we won’t choose the “Assume Linear Model” in the options menu
• Also, x1 & x2 are not specified to be positive, so we don’t check the “Assume Non-negative” box
• If desired, the tolerance may be decreased to 0.1%
Click to edit Master title styleExcel Example
• When we solve the problem, the spreadsheet doesn’t change because our initial guess of x1 = 0 & x2 = 0 is an optimum solution, as we found when we solved the problem analytically.
Click to edit Master title styleExcel Example
However, if we choose initial values of both x1 & x2 as -1, we get the following solution
Click to edit Master title styleConclusions
• So, by varying the initial values, we can get both of the candidate points we found previously
• However, the NLP solver tells us that they are both local minimum points
Click to edit Master title styleReferences
Material for this chapter has been taken from:
• Optimization of Chemical Processes 2nd
Ed.; Edgar, Thomas; David Himmelblau; & Leon Lasdon.