Stochastic Optimal Control!

Transcription

1 Stochastic Control! Robert Stengel! Robotics and Intelligent Systems, MAE 345, Princeton University, 2015 Learning Objectives Overview of the Linear-Quadratic-Gaussian (LQG) Regulator Introduction to Stochastic Robust Control Laws Copyright 2015 by Robert Stengel. All rights reserved. For educational use only. 1 Stochastic Optimal Control! 2

2 Deterministic vs. Stochastic Optimal Control Deterministic control! Known dynamic process precise input precise initial condition precise measurement! Optimal control minimizes J* = J(x*, u*) Stochastic control! Known dynamic process unknown input imprecise initial condition imprecise or incomplete measurement! Optimal control minimizes E{J[x*, u*]} 3 Linear-Quadratic-Gaussian (LQG) Control of a Dynamic Process! z(t) = Hx(t) + n(t) [ ] ˆ!x(t) = F(t)ˆx(t) + G(t)u(t) + K(t) z(t)! Hˆx(t) u(t) =!C(t) ˆx(t) + u command 4

3 Linear-Quadratic (LQ) Control Equations (Continuous-Time Model) Open-Loop System State Dynamics u t = Fx(t) + Gu( t)!x t Control Law =!Cx(t) + u command Closed-Loop System State Dynamics!x ( t) = ( F! GC)x(t) + GCu command Characteristic Equation si! ( F! GC) = 0 How many eigenvalues? n Stable or unstable? Stable, with correct design criteria, F, and G 5 Linear-Quadratic-Gaussian (LQG) Control Equations (Continuous-Time Model) Open-Loop System State Dynamics and Measurement ˆ!x t = Fx(t) + Gu(t) + Lw(t) z( t) = Hx( t) + n( t)!x t State Estimate = Fˆx(t) + Gu(t) + K z( t)! Hˆx ( t) " Control Law u( t) =!Cˆx(t) + u command 6

4 Linear-Quadratic-Gaussian (LQG) Control Equations (Continuous-Time Model) Closed-Loop System State and Estimate Dynamics (neglect command) = Fx(t) + G!Cˆx(t)!x t ˆ!x ( t) = Fˆx(t) + G!Cˆx(t) [ ]+ Lw(t) [ ]+ K " z( t)! Hˆx ( t) How many eigenvalues? 2n Stable or unstable? TBD 7 LQG Separation Property Optimal estimation algorithm does not depend on the optimal control algorithm = P(t)H T N!1 ( t) L T ( t)! P(t)H T N!1 t K t!p(t) = F(t)P(t) + P(t)F T (t) + L( t)w t HP(t) Optimal control algorithm does not depend on the optimal estimation algorithm C t = R!1 t!s ( t) =!Q(t)! F(t) T S( t)! S t G T ( t)s( t) F(t) + S( t)g(t)r!1 (t)g T (t)s t 8

5 LQG Certainty Equivalence Stochastic feedback control is computed from optimal estimate of the state u*(t) =!R!1 G T (t)s(t)ˆx(t) =!C(t) ˆx(t) Stochastic feedback control law is the same as the deterministic control law u*(t) =!R!1 G T (t)s(t)x(t) =!C(t) x(t) 9 Asymptotic Stability of the LQG Regulator! (with no parameter uncertainty)! 10

6 System Equations with LQG Control With perfect knowledge of the system!x ( t) = Fx(t) + G [!Cˆx(t) ]+ Lw(t) ˆ!x ( t) = Fˆx(t) + G!Cˆx(t) [ ]+ K " z( t)! Hˆx ( t) State estimate error!( t)! x( t) " ˆx ( t) State estimate error dynamics!( t) = ( F " KH)! ( t) + Lw t " Kn( t) 11 Control-Loop and Estimator Eigenvalues are Uncoupled "!x t! t ' ' = " ( F ( GC) GC 0 F ( KH " ' ' x t! t ' ' + " L 0 L (K " ' w t n t ' ' Upper-block-triangular stability matrix LQG system is stable because (F GC) is stable (F KH) is stable Estimate error affects state response = F! GC!x t x t + GC" ( t) + Lw( t) Actual state does not affect error response Disturbance affects both equally 12

7 Discrete-Time LQG Controller Kalman filter produces state estimate = "ˆx k!1 ( + )! C k!1ˆx k!1 ( + ) ˆx k! ˆx k ( + ) = ˆx k (!) + K k z k! Hˆx k (!) u k =!C k ˆx k ( + ) " Closed-loop system uses state estimate for feedback control (u command = 0) x k+1 (!) = "x k (!)! C k ˆx k ( + ) 13 Response of Discrete-Time 1 st -Order System to Disturbance, and Kalman Filter Estimate from Noisy Measurement Propagation of Uncertainty Kalman Filter, Uncontrolled System 14

8 Comparison of 1 st -Order Discrete-Time LQ and LQG Control Response Linear-Quadratic Control with Noise-free Measurement Linear-Quadratic-Gaussian Control with Noisy Measurement 15 MATLAB Demo: LQG Rolling Mill Control System Design Example Open- and Closed-Loop Response Maintain desired thickness of shaped beam Account for random! variations in thickness/ hardness of incoming beam! eccentricity in rolling cylinders! measurement errors 16

9 Robust Stochastic Control! 17 Stochastic, Robust, and Adaptive Control Stochastic controller minimize response to random initial conditions, disturbances, and measurement errors perfect knowledge of the plant Robust controller fixed gains and structure minimize likelihood of instability or unsatisfactory performance due to parameter uncertainty in the plant Adaptive controller variable gains and/or structure minimize likelihood of instability or unsatisfactory performance due to plant parameter uncertainty, disturbances, and measurement errors Practical controller may have elements of all three 18

10 Robust Control System Design Make closed-loop response insensitive to plant parameter variations Robust controller! Fixed gains and structure! Minimize likelihood of instability! Minimize likelihood of unsatisfactory performance 19 Probabilistic Robust Control Design Design a fixed-parameter controller for stochastic robustness Monte Carlo Evaluation of competing designs Genetic Algorithm or Simulated Annealing search for best design 20

11 Representations of Uncertainty Characteristic equation of the uncontrolled system si! F = det( si! F)! "(s) = s n + a n!1 s n! a 1 s + a 0 ( s! 2 )(...)( s! n ) = 0 = s! 1 Uncertainty can be expressed in! Elements of F! Coefficients of "(s)! Eigenvalues of F 21 Root Locations for an Uncertain 2 nd -Order System Variation may be represented by! Worst-case, e.g., Upper/lower bounds of uniform distribution! Probability, e.g., Gaussian distribution Uniform Distribution Gaussian Distribution s Plane s Plane 22

12 3-D Stochastic Root Loci for 2 nd -Order Example Root distributions are nonlinear functions of parameter distributions Unbounded parameter distributions always lead to non-zero probability of instability Bounded distributions may be guaranteed to be stable 23 Probability of Satisfying a Design Metric Pr(d,v)! 1 N N i=1 e" C( d), H ( v) Probability of satisfying a design metric! d: Control design parameter vector [e.g., SA, GA, ]! v: Uncertain plant parameter vector [e.g., RNG]! e: Binary indicator, e.g.,! 0: satisfactory 1: unsatisfactory! H(v): Plant! C(d): Controller (Compensator) 24

13 Design Control System to Minimize Probability of Instability Characteristic equation of the closed-loop system! closed"loop (s) = si " F( v) " G( v)c( d) ( s " ' 2 )(...)( s " ' n ) = s " ' 1 closed"loop = 0 Monte Carlo evaluation of probability of instability with uncertain plant parameters Minimize probability of instability using numerical search of control parameters { } min Pr Re! i, i = 1,n d " > 0 25 Control Design Example* u k w y m 1 m 2 Challenge: Design a feedback compensator for a 4 th -order spring-mass system ( the plant ) whose parameters are bounded but unknown! Minimize the likelihood of instability! Satisfy a settling time requirement! Don t use too much control * 1990 American Control Conference Robust Control Benchmark Problem 26

14 Design Cost Function Probability of Instability, Pr i! e i = 1 (unstable) or 0 (stable) Probability of Settling Time Exceedance, Pr ts! e ts = 1 (exceeded) or 0 (not exceeded) Probability of Control Limit Exceedance, Pr u! e u = 1 (exceeded) or 0 (not exceeded) Each metric has a binomial distribution pr( x) = n! k! ( n! k)! p x ( n!k )k " 1! p( x)! ' ( n k ) * + p( x n!k )k " 1! p( x) = probability of exactly k successes in n trials, in 0,1 ~ normal distribution for large n Design Cost Function! High probabilities weighted more than low probabilities! J = apr 2 i + bpr 2 ts + c Pr 2 u! a = 1! b = c = Monte Carlo Evaluation of Probability of Satisfying a Design Metric 1e+09 N Pr k (d,v)! 1 e k C( d), H ( v) N ", k = 1,3 i=1 J = apr i 2 (d,v) + bpr ts 2 (d,v) + cpr u 2 (d,v) Number of Evaluations 1e+07 1e+05 1e+03 Interval Width e+01 1e-06 1e-05 1e Probability or (1 - Probability Binomial Distribution Compute v using random number generators over N trials! Required number of trials depends on outcome probability and desired confidence interval Search for best d using a genetic algorithm to minimize J 28

15 Uncertain Plant* w k u y m 1 m 2 y = x 2 + n Plant dynamic equation! "!x 1!x 2!x 3!x 4! = " ' k m 1 k m k m 2 ' k m 2 0 0! " x 1 x 2 x 3 x 4! + " m 1 0! u + " m 2 w 4 th -Order Plant characteristic equation "!(s) = s 2 s 2 + k m + m 1 2 m 1 m 2 ' = s 2 s 2 2 " n * 1990 American Control Conference Robust Control Benchmark Problem 29 Parameter Uncertainties, Root Locus, and Control Law Parameters of massspring system! Uniform probability density functions for 0.5 < m 1, m 2 < < k < 2 Effects of parameters on root locations (right) x xx x "!(s) = s 2 s 2 + k m 1 + m 2 ' = s 2 s 2 2 m 1 m " n 2 Single-input/single-output feedback control law u( s) =!C ( s)y( s) 30

16 Mass-Spring-Mass Stabilization Requires Compensation Proportional feedback alone cannot stabilize the system Feedback of either sign drives at least one root into the right half plane u( s) =!cy( s) 31 Search-and-Sweep Design of Family of Robust Feedback Compensators Begin with lowest-order feedback compensator C 12 (s) = a 0 + a 1 s b 0 + b 1 s + b 2 s 2! C d Arrange parameters as binary design vector { } d = a 0,a 1,b 0,b 1,b 2 d* = { a 0 *,a 1 *,b 0 *,b 1 *,b 2 *} Search for design vector, d, that minimizes J m 1 = rand 1 m 2 = rand k =!1.5 *rand 1 Monte Carlo evaluation with uncertain parameters, v 32

17 Search-and-Sweep Design of Family of Robust Feedback Compensators 1) Define next higher-order compensator C 22 (s) = a 0 + a 1 s + a 2 s 2 b 0 + b 1 s + b 2 s 2 2) Optimize over all parameters, including optimal coefficients in starting population d = { a 0 *,a 1 *,a 2,b 0 *,b 1 *,b 2 *}! d ** = { a 0 **,a 1 **,a 2 **,b 0 **,b 1 **,b 2 **} 3) Sweep to satisfactory design or no further improvement C 23 (s) = a 0 + a 1 s + a 2 s 2 b 0 + b 1 s + b 2 s 2 + b 3 s 3 C 33 (s) = a 0 + a 1 s + a 2 s 2 + a 3 s 3 b 0 + b 1 s + b 2 s 2 + b 3 s 3 C 34 (s) = a 0 + a 1 s + a 2 s 2 + a 3 s 3 b 0 + b 1 s + b 2 s 2 + b 3 s 3 + b 4 s Design Cost and Probabilities for Optimal 2 nd - to 5 th -Order Compensators Number of Zeros = Number of Poles nd Order 3rd Order 4th Order 5th Order Cost Prob Instab Prob Ts Viol x 0.01 Prob Cont Viol x

18 Next Time:! Parameter Estimation and Adaptive Control! 35 Supplemental Material 36

19 Linear-Quadratic Gaussian Optimal Control Law Minimize expected value of cost, subject to uncertainty* min E(J)! E(J*) u Stochastic optimal feedback control law combines the linear-optimal control law with a linear-optimal state estimate u*(t) =!R!1 G T (t)s(t)ˆx(t) =!C(t) ˆx(t)! where ˆx(t) is an optimal estimate of the state perturbation * See reading for details 37 Example: Probability of Stable Control of an Unstable Plant X-29 Aircraft Longitudinal dynamics for a Forward-Swept-Wing Demonstrator F =!2gf 11 / V "V 2 f 12 / 2 "Vf 13!g!45 / V 2 "Vf 22 / "V 2 f 32 / 2 "Vf ( ( (; G = ( ( ' g 11 g g 31 g V ( ( ; x = ( ( q ( ( '( * ( ' Nominal eigenvalues (one unstable)! 1" 4 = "0.1 ± j, " 5.15, 3.35 Air density and airspeed,! and V, have uniform distributions(±30) 10 coefficients have Gaussian distributions (! = 30) p = "! V f 11 f 12 f 13 f 22 f 32 f 33 g 11 g 12 g 31 g 32 T Environment Uncontrolled Dynamics Control Effect 38

20 LQ Regulators for the Example Three stabilizing feedback control laws Case a) LQR with low control weighting Case b) LQR with high control weighting Q = diag( 1,1,1,0 ); R = ( 1,1);! 1" 4nominal = 35, 5.1, 3.3,.02 C = "11 "3 "1.1 Q = diag( 1,1,1,0 ); R = ( 1000,1000);! 1" 4nominal = "5.2,"3.4,"1.1,".02 C = " "63 "16 "1.9 ( ' ( ' Case c) Case b with gains multiplied by 5 for bandwidth (loop-transfer) recovery! 1" 4nominal = "32, 5.2, 3.4, 0.01 C = " "313 "81 "1.1" 9.5 ( ' 39 Stochastic Robustness! (Ray, Stengel, 1991) "! Distribution of closed-loop roots with "! Gaussian uncertainty in 10 parameters "! Uniform uncertainty in velocity and air density "! 25,000 Monte Carlo evaluations "! Probability of instability "! a) Pr = "! b) Pr = "! c) Pr = Stochastic Root Locus 40

21 Probabilities of Instability for the Three Cases 41 Case a: Low LQ Control Weights Stochastic Root Loci for the Three Cases with Gaussian Aerodynamic Uncertainty! Case b: High LQ Control Weights Probabilities of instability with 30 uniform aerodynamic uncertainty!! Case a: 3.4 x 10-4!! Case b: 0!! Case c: 0! Case c: Bandwidth Recovery 42