outline multiparametric linear programming with
TRANSCRIPT
Automatic Control Laboratory, ETH Zürichwww.control.ethz.ch
Multiparametric Linear Programmingwith
Applications to Control
Manfred Morari
Colin Jones, Miroslav Baric, Melanie Zeilinger
Outline
1. Motivation
2. Geometry of Parametric Programming
3. Approximate Controllers
4. Industrial Applications
Motivation
ControllerComplexity
ControlPerformance
Motivation
Why is complexity important?
• Design / maintenance effort– Process industries
• Implementation cost (hardware)– Automotive– Electronics– Electrical
DC/DC Converterfor Mobile Phones
ComplexityPe
rfor
man
ce
Current
Target Performance
MPC
Goal: Minimize complexity for given performance
Complexity=
Grams silicon=
Torque Control of anInduction Machine
Complexity
Perf
orm
ance
CurrentAvailableProcessor
Goal: Maximise performance for specific processor
Specific Example:Optimal Control Synthesis
Given a performance index
u = f(x)
Plantcontrolaction u
plantstate x
Compute control action
Obtain U*(x)
Plant output y
plant state xcontrol u0*
Receding Horizon ControlOn-Line Optimization
Optimization Problem
Model Predictive Control (MPC)
ParametricOptimization
Explicit Solution
Plant output y
plant state xcontrol u*
(=Look-Up Table)
Receding Horizon ControlOff-Line Optimization
off-line
Seron, De Doná and Goodwin, 2000Johansen, Peterson and Slupphaug, 2000Bemporad, Morari, Dua and Pistokopoulos, 2000
Explicit MPC
Solution ofBellman equation
Explicit MPC
Explicit MPC Controller Properties
Explicit MPC control law = MPC control law
Guarantees using MPC design techniques• Stability• Feasibility• Invariance
Pros– Easy to implement– Fast on-line evaluation– Analysis of closed-loop system possible
Challenges– Number of controller regions can become large– Computation time may become prohibitive– Numerics
Multi-parametric controllers Online vs. Offline Complexity
Com
puta
tion
Tim
e
Complexity ofonline optimization
log(Problem Complexity)Complexity ofexplicit controller
Optimality Trade-off
Complexity
Perf
orm
ance
Explicit
MPC
Goal: Trade-off performance and complexity
Research Directions1. Reduce complexity of online optimization
– Rao, Wright, Rawlings, 1998– Diehl, Björnberg, 2004– Wang, Boyd, 2007
2. Reduce complexity of explicit solution(i.e., number of regions)– Jones, Baric, Morari, 2007– Lincoln, Rantzer, 2006– Bemporad, Filippi, 2004– Johansen, Grancharova, 2003
3. Combination of 1. and 2.– Panocchia, Rawlings, Wright, 2006– This presentation
Outlook
Complexity
Perf
orm
ance
Explicit
Goal: Trade-off performance and complexity
Outline
1. Motivation
2. Geometry of Parametric Programming
3. Approximate Controllers
4. Industrial Applications
MPC → Parametric Programming
Primal
p = 1,1
MPC → Parametric Programming
Primal
p = 1,1
Parameter = state
Geometry of mpLP : PrimalPrimal
u
x
P
Geometry of mpLP : PrimalPrimal
u
x
P
Feasible inputs for givenstate x0
x0
P
Geometry of mpLP : PrimalPrimal
u
x
J
c0u
x0
Cost of applying inputs for givenstate x0
Geometry of mpLP : PrimalPrimal
x
J
Optimal cost and input forstate x0
x0
P
u
Goal : Minimize J
P
Geometry of mpLP : PrimalPrimal
u
x
J
Cost/control affine in each regionCompute optimal input for each state
Goal: Compute facetsof cost function
Parametric = Facet/Vertex EnumerationPrimal Dual
Facet Enumeration ofImplicit polyhedron
Vertex Enumeration ofImplicit polyhedron
Current pLP AlgorithmsGoal: Compute the vertices (facets) of implicitly
defined polyhedronBased on classic vertex/facet enumeration
Primal : Gift Wrapping Dual : Pivoting
Current pLP AlgorithmsGoal: Compute the vertices (facets) of implicitly
defined polyhedronBased on classic vertex/facet enumeration
Primal : Gift Wrapping Dual : Pivoting
Current pLP AlgorithmsGoal: Compute the vertices (facets) of implicitly
defined polyhedronBased on classic vertex/facet enumeration
Primal : Gift Wrapping Dual : Pivoting
Current pLP AlgorithmsGoal: Compute the vertices (facets) of implicitly
defined polyhedronBased on classic vertex/facet enumeration
Primal : Gift Wrapping Dual : Pivoting
Current pLP AlgorithmsGoal: Compute the vertices (facets) of implicitly
defined polyhedronBased on classic vertex/facet enumeration
Primal : Gift Wrapping Dual : Pivoting
mpLP in ControlNumber of facets/regions…
– online computation time– storage– offline processing time
… can grow very quickly with problem size
Current algorithms are non-incremental– output meaningless until complete– cannot stop early
Outline
1. Motivation
2. Geometry of Parametric Programming
3. Approximate Controllers
4. Industrial Applications
Approximation as Separation
J*(x)
Goal: Compute ‘simpler’ approximation of J*Without computing J¤!
Approximation as SeparationFind function such that :• Stability :• Invariance :• Performance :
J*(x)
Related work : Lincoln and Rantzer 2006
Approximation as SeparationFind function such that :• Stability :• Invariance :• Performance :
All separatingfunctions are
stabilizing
Level setsare invariant
Polyhedral SeparationGiven polyhedra S µ P find minimum complexitypolyhedron Q such that S µ Q µ P
P
Q
S
Polyhedral SeparationApplications:1. Geometric surface approximation2. Multi-resolution transmission3. Polyhedral function approximation4. … P
Q
S
Classic method for vertex enumeration
Rediscovered many (many) times
Goal: Compute the convex hull of a set of points
Idea: Build convex hull iteratively by addingpoints one at a time
Beneath/Beyond
Fourier1824
Motzkin1953
Kallay1981
Jones, Baric, Morari, 2007
Parametric Linear Optimisation
How to recover the approximate control law?
Barycentric Interpolation
Weight vertices based onintegral of the surface area ofthe polar dual (Stokes Law).
Goal: Given set of vertices V andfunction f, compute a smoothfunction g such that
1. g(v) = f(v) for all v in V2. Domain of g = conv(V)3. (v,g(v)) 2 conv({vi,f(vi)})
Schaefer, 2007
Beneath/Beyond in ControlBenefits1. Constructive procedure2. Stop at a given complexity (very nice)3. Stabilizing and invariant4. Tradeoff complexity with region of attraction
and/or performance5. Optimal solution in finite time
Jones, Baric and Morari, 2007
Approximate Solutions
• Random system• 808 regions
-8 -6 -4 -2 0 2 4 6 8 10-2
0
2
4
6
8
10
Approximate Solutions
• 808 40 regions• Stabilizing• Small performance loss
-8 -6 -4 -2 0 2 4 6 8 10-2
0
2
4
6
8
10
Approximate Solutions: Example
12128 357 regions
Beneath/Beyond in ControlLimitations1. Approximation >> optimal (exponential)2. Compute vertices and halfspaces (exponential)3. B/B dislikes degeneracy (exponential)4. Greedy approximation (not optimal)
Great for some, rather less so for others!
Jones, Baric and Morari, 2007
Extensions of theory and computationControl :• System type
– Switched linear / piecewise affine– Uncertainty in inputs and/or parameters
• Performance Objective– Quadratic– Min / max– Infinite horizon
• Mixed online/offline computation– Warm start with online complexity guarantees– Continuous trade-off between online/offline effort
Optimal tradeoff
Perf
orm
ance
Arithmetic Operations
Combination of online and offline methods to trade-offperformance and complexity
Onlineoptimization
(Approximate)Explicit solution
Online optimizationwith warmstart
Online/Offline Tradeoff: ExampleOnline Optimization:
0 0.5 1 1.5 2 2.5
x 106
0
1
2
3
4
# arithmetic operations
Stability Bounderro
r
Phase 1
Simplex
Worst case assessed by sampling
0 5 10 15
x 105
0
1
2
3
4
erro
r
# arithmetic operations
Stability Bound
Simplex
Beneath/Beyond
extrapolated
practically solvable
Online/Offline Tradeoff: ExampleApproximate Explicit Solution:
Worst case assessed by sampling
0 5 10 15
x 105
0
1
2
3
4
erro
r
# arithmetic operations
Stability Bound
Simplex
Beneath/Beyond
Online/Offline Tradeoff: ExampleOnline optimization starting from approximate explicitsolution:
Active Set
optimal
Combination between online and offline solutionoutperforms explicit solution
Worst case assessed by sampling
0 5 10 15
x 105
0
1
2
3
4
erro
r
# arithmetic operations
Stability Bound
Simplex
Beneath/Beyond
Online/Offline Tradeoff: ExampleOnline optimization starting from approximate explicitsolution:
Active Set
0 1 2 3 4 5 6 7
x 105
0
0.2
0.4
0.6
0.8
1
erro
r
# arithmetic operations
Stabilizing region
offline
online/
offline
offline online/offline
0 0.2 0.4 0.6 0.8 110
3
104
105
106
107
θ=online/offline tradeoff
optimal
# ar
ithm
etic
ope
ratio
ns
Online Offline
Tradeoff map: Example
Look for the minimum
Get best tradeoff
0 0.2 0.4 0.6 0.8 110
3
104
105
106
107
optimal
# ar
ithm
etic
ope
ratio
ns
Online Offlineθ=online/offline tradeoff
Tradeoff map: Example
errorlevel curves
Best tradeoff
Parametric Programming and ControlParametric linear programming (many, many more)1950s : Orchard-Hays, …1970s : Klatte, Kummer, …1990s- : Baotić, Barić, Bemporad, Borelli, Filippi, Gal,
Johansen, Jones, Murty, Spøtvold, Terlaky, Tøndel, …
Explicit MPC1990s- : Baotić, Barić, Bemporad, De Dona, Dua, Dumur,
Filippi, Goodwin, Grieder, Johansen, Jones, Kerrigan, Kvasnica,Lincoln, Löfberg, Maciejowski, Mayne, Morari, Olaru, Parrilo,Pistikopulous, Raković, Rantzer, Sakizlis, Schroeder, Seron,Tøndel, …
• All results were obtained with the MPT toolbox
http://control.ethz.ch/~mpt
• MPT is a MATLAB toolbox that provides efficientcode for– (Non)-Convex Polytopic Manipulation– Multi-Parametric Programming– Control of PWA and LTI systems
Conclusions
MPT in the World
10000+ downloads in 5 years
Outline
1. Motivation
2. Geometry of Parametric Programming– Focus on linear
3. Approximate Controllers
4. Industrial Applications
Switch-mode DC-DC Converter
d
unregulated DC voltage low-pass filter load
dually operated switches
Switched circuit: supplies power to load with constant DC voltage
Illustrating example: synchronous step-down DC-DC converter
regulated DC voltage
S1
S2
Operation Principle
k-th switching period
duty cycle
• Length of switching period Ts constant (fixed switching frequency)
• Switch-on transition at kTs, k2N • Switch-off transition at (k+d(k))Ts (variable pulse width)
• Duty cycle d(k) is real variable bounded by 0 and 1
S1 = 1S2 = 0
S1 = 0S2 = 1
kTs (k+1)Ts(k+d(k))Ts
S1 = 1S2 = 0
S1 = 0S2 = 1
kTs (k+1)Ts(k+d(k))Ts
duty cycle d(k)
Mode 1 and 2
mode 1: mode 2: d
S1
S2
Control Objective
unregulated DC input voltagedisturbance
duty cyclemanipulated variable
regulated DC output voltagecontrolled variable
inductor currentstate
capacitor voltagestate
Regulate DC output voltage by appropriate choice of duty cycle
Control Objectives
Minimize (average) output voltage error andchanges in duty cycle
Respect constraint on current limit
Translate in Receding Horizon Control (RHC) problem
State-feedback Controller:Polyhedral Partition
Colors correspond to the 121 polyhedra
PWA state-feedbackcontrol law:
computed in 100s using theMPT toolbox
121 polyhedra after simplificationwith optimal merging
Start up: Experiment (1/2)
Inductor current
Output voltage
330mA
2V
Time1msec
6V
~1A
660mA
Input voltage @10V, Output voltage reference @6V, Current limit @1A
Experiments at U. Sannio: Papafotiou, Frasca, Vasca, Borrelli
Start up: Experiment (2/2)
Time1msec
~1A
660mA
Input voltage @10V, Output voltage reference @6V, Current limit @1A
Duty Cycle1
05V
Inductor current
330mA
Experiments at U. Sannio: Papafotiou, Frasca, Vasca, Borrelli
Direct Torque Control
Physical Setup:• Three-level DC link inverter
driving a three-phase symmetricinduction motor
• Binary control inputs
Control Objectives:• Keep torque, stator flux and neutral point potential
within given bounds• Minimize average switching frequency (losses)
Reduction of switching frequency by up to 45 % (in average 25 %)with respect to ABB’s commercial DTC scheme (ACS 6000)
Physical Setup:• Cascade of hydroelectric power plants• Turbines and weirs (adjustable water flow) • Lock operation disturbances
Control Objectives:• Respect constraints on concession level• Minimize flow changes at weirs and turbines• Maximize income from power generation
River Power Plants
Optimal coordination between plants leads to 1.) Major damping of upstream water flow disturbances; 2.) Strict respect of concessions level constraints
Electronic Throttle ControlPhysical Setup:• Valve (driven by DC motor) regulates air inflow to the car engine• Friction nonlinearity• Limp-Home nonlinearity• Physical constraints
Control Objectives:• Minimize steady-state regulation error• Achieve fast transient behavior without overshoot
Systematic controller synthesis procedure. On average twice asfast transient behavior compared to state-of-the-art PID
controller with ad-hoc precompensation of nonlinearities.
Traction ControlPhysical Setup:• Improve driver's ability to control vehicle under adverse external conditions (wet or icy roads)• Tire torque is nonlinear function of slip• Uncertainties and constraints
Control Objectives:• Maximize tire torque by keeping tire
slip close to the desired value
Experimental results: 2000 Ford Focus on a Polished Ice Surface;Receding Horizon controller with 20 ms sampling time
Tire Slip
Tire
Tor
que
Piecewise affineapproximation
Adaptive Cruise ControlPhysical Setup:
Control Objectives:• Track reference speed• Respect traffic rules• Consider all objects on all lanes
Optimal state-feedback control law successfully implemented andtested on a research car Mercedes E430 with 80ms sampling time
Longitudinal and LateralControl
Sensors: IRScanner,Cameras, Radar
TestVehicle
Traffic Scene
Virtual TrafficScene
Smart Damping Materials
– Device suppresses vibration– External power source for operation is not required– Weight and size of the device have to be kept to a
minimum
• Demands
• Idea– Switched Piezoelectric (PZT) Patches
• Problem– What is the optimal switching law for optimal vibration suppression?
PZTPZT
Niederberger, Moheimani
Control of Anaesthesia
Physical Setup:• Patient undergoing surgery• Analgesic infusion pump
Control Objectives:• Minimize stress reaction to
surgical stimulation(by controlling mean arterial pressure)
• Minimize drug consumption
Excellent performance of administration scheme,mean arterial pressure variations kept within bounds
Control of Thermal Print-HeadsPhysical Setup:• Thermal print-head
with ~ 1400 heat elements• Binary control inputs• Printing on a wide range
of materials
Control Objectives:• Maximize printout quality• Achieve robustness to parameter variations
90% quality gain over traditional controllers [ANSI X3.182-1990];Straight-forward design method for print-head controller
Conclusions
• Control synthesis tools emerging fromconfluence of mathematical programming,computational geometry and control theory
• Tools for constrained switched linear systemsaffecting industrial applications
• Performance/complexity trade-off importantfor cost sensitive application domains arebeing addressed