Time-Invariant Linear Quadratic Regulators!

Similar documents
Time-Invariant Linear Quadratic Regulators Robert Stengel Optimal Control and Estimation MAE 546 Princeton University, 2015

Stochastic and Adaptive Optimal Control

First-Order Low-Pass Filter!

Return Difference Function and Closed-Loop Roots Single-Input/Single-Output Control Systems

Dynamic Optimal Control!

First-Order Low-Pass Filter

Suppose that we have a specific single stage dynamic system governed by the following equation:

6.241 Dynamic Systems and Control

Stochastic Optimal Control!

OPTIMAL CONTROL AND ESTIMATION

LQR, Kalman Filter, and LQG. Postgraduate Course, M.Sc. Electrical Engineering Department College of Engineering University of Salahaddin

Time Response of Dynamic Systems! Multi-Dimensional Trajectories Position, velocity, and acceleration are vectors

Lecture 2: Discrete-time Linear Quadratic Optimal Control

7.1 Linear Systems Stability Consider the Continuous-Time (CT) Linear Time-Invariant (LTI) system

UCLA Chemical Engineering. Process & Control Systems Engineering Laboratory

Topic # Feedback Control Systems

Robotics. Control Theory. Marc Toussaint U Stuttgart

Optimal Control. Quadratic Functions. Single variable quadratic function: Multi-variable quadratic function:

1. Find the solution of the following uncontrolled linear system. 2 α 1 1

Quadratic Stability of Dynamical Systems. Raktim Bhattacharya Aerospace Engineering, Texas A&M University

OPTIMAL CONTROL. Sadegh Bolouki. Lecture slides for ECE 515. University of Illinois, Urbana-Champaign. Fall S. Bolouki (UIUC) 1 / 28

Linear-Quadratic-Gaussian Controllers!

Linear Algebra. P R E R E Q U I S I T E S A S S E S S M E N T Ahmad F. Taha August 24, 2015

4F3 - Predictive Control

Contents. 1 State-Space Linear Systems 5. 2 Linearization Causality, Time Invariance, and Linearity 31

Control Systems! Copyright 2017 by Robert Stengel. All rights reserved. For educational use only.

Homework Solution # 3

MODERN CONTROL DESIGN

Theory in Model Predictive Control :" Constraint Satisfaction and Stability!

EE C128 / ME C134 Feedback Control Systems

EL2520 Control Theory and Practice

Optimal Control and Estimation MAE 546, Princeton University Robert Stengel, Preliminaries!

and the nite horizon cost index with the nite terminal weighting matrix F > : N?1 X J(z r ; u; w) = [z(n)? z r (N)] T F [z(n)? z r (N)] + t= [kz? z r

State Regulator. Advanced Control. design of controllers using pole placement and LQ design rules

Steady State Kalman Filter

Linear Quadratic Zero-Sum Two-Person Differential Games Pierre Bernhard June 15, 2013

Lyapunov Stability Theory

Problem 1 Cost of an Infinite Horizon LQR

Topic # /31 Feedback Control Systems. Analysis of Nonlinear Systems Lyapunov Stability Analysis

Chap 8. State Feedback and State Estimators

MS-E2133 Systems Analysis Laboratory II Assignment 2 Control of thermal power plant

MATH4406 (Control Theory) Unit 6: The Linear Quadratic Regulator (LQR) and Model Predictive Control (MPC) Prepared by Yoni Nazarathy, Artem

Problem Set 3: Solution Due on Mon. 7 th Oct. in class. Fall 2013

Video 6.1 Vijay Kumar and Ani Hsieh

Lecture 9: Discrete-Time Linear Quadratic Regulator Finite-Horizon Case

Here represents the impulse (or delta) function. is an diagonal matrix of intensities, and is an diagonal matrix of intensities.

CDS 110b: Lecture 2-1 Linear Quadratic Regulators

May 9, 2014 MATH 408 MIDTERM EXAM OUTLINE. Sample Questions

Linear Quadratic Regulator (LQR) Design I

EML5311 Lyapunov Stability & Robust Control Design

SUCCESSIVE POLE SHIFTING USING SAMPLED-DATA LQ REGULATORS. Sigeru Omatu

4F3 - Predictive Control

CDS 101/110: Lecture 3.1 Linear Systems

ESC794: Special Topics: Model Predictive Control

Linearized Equations of Motion!

MATH 235: Inner Product Spaces, Assignment 7

Chap 4. State-Space Solutions and

EE363 homework 2 solutions

only nite eigenvalues. This is an extension of earlier results from [2]. Then we concentrate on the Riccati equation appearing in H 2 and linear quadr

Optimization-Based Control

MCE693/793: Analysis and Control of Nonlinear Systems

Optimization of Linear Systems of Constrained Configuration

Appendix A Solving Linear Matrix Inequality (LMI) Problems

Lyapunov Stability Analysis: Open Loop

Mathematical Systems Theory: Advanced Course Exercise Session 5. 1 Accessibility of a nonlinear system

Decentralized control with input saturation

MCE693/793: Analysis and Control of Nonlinear Systems

INVERSE MODEL APPROACH TO DISTURBANCE REJECTION AND DECOUPLING CONTROLLER DESIGN. Leonid Lyubchyk

UCLA Chemical Engineering. Process & Control Systems Engineering Laboratory

Linear Matrix Inequality (LMI)

Nonlinear Control. Nonlinear Control Lecture # 3 Stability of Equilibrium Points

MCE/EEC 647/747: Robot Dynamics and Control. Lecture 8: Basic Lyapunov Stability Theory

On a Nonlinear Riccati Matrix Eigenproblem

Chapter III. Stability of Linear Systems

9 Controller Discretization

Outline. 1 Linear Quadratic Problem. 2 Constraints. 3 Dynamic Programming Solution. 4 The Infinite Horizon LQ Problem.

Linear State Feedback Controller Design

EE C128 / ME C134 Final Exam Fall 2014

Problem Description The problem we consider is stabilization of a single-input multiple-state system with simultaneous magnitude and rate saturations,

Computational Issues in Nonlinear Dynamics and Control

8. Diagonalization.

1.2. Direction Fields: Graphical Representation of the ODE and its Solution Let us consider a first order differential equation of the form dy

2 The Linear Quadratic Regulator (LQR)

Digital Control Engineering Analysis and Design

Exam. 135 minutes, 15 minutes reading time

Linear Quadratic Zero-Sum Two-Person Differential Games

Robotics: Science & Systems [Topic 6: Control] Prof. Sethu Vijayakumar Course webpage:

1 Lyapunov theory of stability

Subject: Optimal Control Assignment-1 (Related to Lecture notes 1-10)

CALIFORNIA INSTITUTE OF TECHNOLOGY Control and Dynamical Systems. CDS 110b

Lecture 5 Linear Quadratic Stochastic Control

ẋ n = f n (x 1,...,x n,u 1,...,u m ) (5) y 1 = g 1 (x 1,...,x n,u 1,...,u m ) (6) y p = g p (x 1,...,x n,u 1,...,u m ) (7)

Linear Systems. Manfred Morari Melanie Zeilinger. Institut für Automatik, ETH Zürich Institute for Dynamic Systems and Control, ETH Zürich

Linear-Quadratic Control System Design

Automatic Control II Computer exercise 3. LQG Design

Module 02 CPS Background: Linear Systems Preliminaries

Lecture 1 From Continuous-Time to Discrete-Time

IEOR 265 Lecture 14 (Robust) Linear Tube MPC

Automatic Control 2. Nonlinear systems. Prof. Alberto Bemporad. University of Trento. Academic year

The Newton-ADI Method for Large-Scale Algebraic Riccati Equations. Peter Benner.

Transcription:

Time-Invariant Linear Quadratic Regulators Robert Stengel Optimal Control and Estimation MAE 546 Princeton University, 17 Asymptotic approach from time-varying to constant gains Elimination of cross weighting in cost function Controllability and observability of an LTI system Requirements for closed-loop stability Algebraic Riccati equation Equilibrium response to commands Copyright 17 by Robert Stengel. All rights reserved. For educational use only. http://www.princeton.edu/~stengel/mae546.html http://www.princeton.edu/~stengel/optconest.html 1 Continuous-Time, Linear, Time-Invariant System Model x = Fx + Gu + Lw, x(t o ) given y = H x x + H u u + H w w Comment: (.) notation distinguishes linear-system variables from nonlinear-system variables

Linear-Quadratic Regulator: Finite Final Time x = Fx + Gu u = R 1 M T + G T P t x t = C( t)x( t) + 1 t ) f + * (,+ J = 1 xt (t f )P(t f )x(t f ) x T u T Q M M T R x u -+ dt. /+ P ( t) = F GR 1 M T T P( t) P( t) F GR 1 M T + P( t)gr 1 G T P( t) + MR 1 M T Q = P f P t f 3 Transformation of Variables to Eliminate Cost Function Cross Weighting Original LTI minimization problem min u 1 J 1 = 1 t f x T 1 Q 1 x 1 + x T 1 M 1 u 1 + u 1 R 1 u 1 dt subject to x 1 = F 1 x 1 + G 1 u 1 Can we find a transformation such that min u J = 1 t f x T Q x + u T R u dt subject to x = F x + G u = min u 1 J 1 4

Artful Manipulation Rewrite integrand of J 1 to eliminate cross weighting of state and control x T 1 Q 1 x 1 + x T 1 M 1 u 1 + u 1 R 1 u 1 = x T 1 Q 1 M 1 R 1 T ( 1 M 1 )x 1 + u 1 + R 1 1 M T T 1 x 1 R1 u 1 + R 1 1 M T 1 x 1 x 1 T Q x 1 + u T R 1 u The transformation produces the following equivalences x = x 1 u = u 1 + R 1 1 M 1 T x 1 Q = Q 1 M 1 R 1 1 M 1 T R = R 1 5 (Q,R) and (Q,M,R) LQ Problems are Equivalent x = x 1 x = x 1 u = u 1 + R 1 1 M 1 T x 1 Q = Q 1 M 1 R 1 T 1 M 1 R = R 1 x = F x + G u x = F x 1 + G u 1 + R 1 1 M T 1 x 1 = F + R 1 T ( 1 M 1 )x 1 + G u 1 = x 1 = F 1 x 1 + G 1 u 1 G = G 1 F = F 1 G R 1 T 1 M 1 = F 1 G 1 R 1 1 M 1 T Therefore, the forms are equivalent Whatever we prove for a (Q,R) cost function pertains to a (Q,M,R) cost function 6

x = x + u; x = 1 p ( t) = 1 p( t) + p ( t) = 1 p t f Recall: LQ Optimal Control of an Unstable First-Order System f = 1; g = 1 Control gain = p( t) u = p( t)x x = 1 p( t) x 7 Riccati Solution and Control Gain for Open-Loop Stable and Unstable 1 st -Order Systems P( t f ) = Variations in control gains are significant only in the last 1- of the illustrated time interval As time interval increases, percentage decreases 8

P() Approaches Steady State as t f -> P With M =, = Q F T P( t) P t t f P( t f ) = { F + P( t)gr 1 G T P( t) }dt from t f to Progression of initial Riccati matrix is monotonic with increasing final time for t f > t f1, J *( t f ) J *( t f1 ), P P 1 (see eq. 5.4-9 to 5.4-11, OCE) Rate of change approaches zero with increasing final time dp dt t f 9 Algebraic Riccati Equation and Constant Control Gain Matrix Steady-state Riccati solution (M = ) Q F T P P F + PGR 1 G T P Q F T P SS P SS F + P SS GR 1 G T P SS = Steady-state control gain matrix C ss = R 1 G T P t f = R 1 G T P ss = 1

Controllability of a LTI System Controllability: All elements of the state can be brought from arbitrary initial conditions to zero in finite time x = Fx + Gu x() = x ; x(t finite ) = System is Completely Controllable if Controllability Matrix = G FG F n1 G n nm has Rank n 11 Controllability Examples For non-zero coefficients F = G 1 n n ); G = () n FG ( = n ) 3 n n ) * Rank = ( ) () F = G FG 1 n n ( = n 4 n ); G = () n ) () ) ) * Rank = ( G F = 1 b ; G = b FG = b Rank = 1 G F = FG 1 b = b b b ; G = b Rank = 1

Requirements for Guaranteed Closed-Loop Stability 13 Optimal Cost with Feedback Control = 1 J *( t f ) = 1 t f With u( t) = C( t)x = R 1 G T P( t)x t f With terminal cost = x * T Qx * + u* T Ru* dt Substitute optimal control law in cost function x * T T Qx * + C( t)x * R C( t)x * dt = 1 t f x * T Qx * + x * T ( t)c T ( t)rc( t)x *( t) dt 14

J *( t f ) = 1 Optimal Cost with LQ Feedback Control t f Consolidate terms x * T Q + C T ( t)rc( t) x * t From eq. 5.4-9, OCE, optimal cost depends only on the initial state and J ( t f ) = 1 x *T ()Px *() dt 15 Optimal Quadratic Cost Function is Bounded J *() = lim J *( t f ) = 1 t f 1 t f ( t f x * T Q + C T ( t)rc( t) x *( t) dt x * T Q + C T ( t)rc( t) x *( t) dt 1 x * T Q + C T RC ( x * t dt = 1 x *T ()Px *() As final time goes to infinity J is bounded and positive provided that Because J is bounded, C is a stabilizing gain matrix Q > R > 16

Requirements for Guaranteeing Stability of the LQ Regulator x = Fx + Gu = [ F GC]x Closed-loop system is stable whether or not open-loop system is stable if... Q > R >... and (F,G) is a controllable pair Rank G FG F n1 G = n 17 Lyapunov Stability of the LQ Regulator x = [ F GC]x = F GR 1 G T P x V = x T ( t)px ( t) + x T ( t)px( t) = x T t Lyapunov function V x( t) = ( xt t)px t + F GR 1 G T T { P P } x t P F GR 1 G T P Rate of change of Lyapunov function 18

Lyapunov Stability of the LQ Regulator V = x T t Algebraic Riccati equation Q F T P PF + PGR 1 G T P = Substituting in rate equation + F GR 1 G T T { P P } x t P F GR 1 G T P = x T ( t) { Q + PGR 1 G T P}x( t) Defining matrix is positive definite Therefore, closed-loop system is stable 19 Less Restrictive Stability Requirements Q may be positive semi-definite if (F,D) is an observable pair, where Q D T D, where D may not be ( n n) Observability requirement Rank D T F T D T F T n1 D T = n

Observability Example x 1 x = 1 () n (*) n y = 1 x 1 x x 1 x = Hx t = Fx t For non-zero coefficients H T F T H T = ( n 1 ( n ) ) * Rank = 1 Even Less Restrictive Stability Requirements If F contains stable modes, closed-loop stability is guaranteed if (F,G) is a stabilizable pair (F,D) is a detectable pair

Stability Requirements with Cross Weighting, M, in Cost Function If F contains stable modes, closed-loop stability is guaranteed if [(F GR -1 M T ),G] is a stabilizable pair [(F GR -1 M T ),D] is a detectable pair (Q GR -1 M T ) R > 3 Example: LQ Optimal Control of a First-Order LTI System Cost Function J = 1 ( )x (t f ) + lim 1 t f t f t o ( qx + ru )dt Open-Loop System x = f x + gu Algebraic Riccati Equation q fp + g p r = p fr g p qr g = u = gp r x = cx Choose positive solution of p = fr g ± Control Law fr g + qr g = fr 1± 1+ g ) g ) fr ( * qr,, + 4

Example: LQ Optimal Control of a First-Order LTI System Closed-Loop System x = f g p r ( x = ( f c)x Stability requires that ( f c) < If f <, then system is stable with no control ( c = ) 5 Example: LQ Optimal Control of a First-Order LTI System If f > (unstable), and r >, then fr >, and g p = fr 1+ 1+ g ) g ) fr ( * qr,, + If q, and g, then p fr 1+ 1 q g ( = fr g and closed-loop system is, as q, f g r p ( = f g r fr g ( = f f = f Stable closed - loop system is mirror image of unstable open - loop system when q = 6

Solution of the Algebraic Riccati Equation 7 Solution Methods for the Continuous- Time Algebraic Riccati Equation Q F T P PF + PGR 1 G T P = 1) Integrate Riccati differential equation to steady state ) Explicit scalar equations for elements of P a) Difficult for n > 3 b) May use symbolic math (MATLAB Symbolic Math Toolbox, Mathematica,...) 8

Example: Scalar Solution for the Algebraic Riccati Equation Q F T P PF + PGR 1 G T P = Second-order example q 11 q f 11 f 1 f 1 f T p 11 p 1 g 11 g 1 + p 1 p g 1 g p 11 p 1 p 1 p p 11 p 1 p 1 p r 11 r 1 g 11 g 1 g 1 g T f 11 f 1 f 1 f p 11 p 1 p 1 p = Solve three scalar equations for p 11, p 1, and p 9 More Solutions for the Algebraic Riccati Equation Q F T P PF + PGR 1 G T P = See OCE, Section 6.1 for Kalman-Englar method Kleinman s method MacFarlane-Potter method Laub s method [used in MATLAB] 3

Equilibrium Response to a Command Input 31 Steady-State Response to Commands x = Fx + Gu + Lw, x(t o ) given y = H x x + H u u + H w w State equilibrium with constant inputs... = Fx *+Gu*+Lw * x* = F 1 Gu*+Lw *... constrained by requirement to satisfy command input y* = H x x * +H u u * +H w w * 3

Steady-State Response to Commands Equilibrium that satisfies a commanded input, y C = Fx *+Gu*+Lw * y* = H x x *+H u u*+h w w * Combine equations y C = F H x G H u x * u * + L H w w * (n + r) x (n + m) 33 Equilibrium Values of State and Control to Satisfy Commanded Input Equilibrium that satisfies a commanded input, y C x * u* = F H x G H u (1 (Lw * A (1 y C ( H w w * (Lw * y C ( H w w * A must be square for inverse to exist Then, number of commands = number of controls 34

Inverse of the Matrix F H x G H u 1 A 1 = B = B 11 B 1 B 1 B x * u* = B 11 B 1 B 1 B B ij have same dimensions as equivalent blocks of A Equilibrium that satisfies a commanded input, y C (Lw * y C ( H w w * x* = B 11 Lw * +B 1 y C H w w * u* = B 1 Lw * +B y C H w w * 35 Elements of Matrix Inverse and Solutions for Open-Loop Equilibrium Substitution and elimination (see Supplement) B 11 B 1 B 1 B = F 1 (GB 1 + I n ) F 1 GB B H x F 1 (H x F 1 G + H u ) 1 Solve for B, then B 1 and B 1, then B 11 x* = B 1 y C ( B 11 L + B 1 H w )w * u* = B y C ( B 1 L + B H w )w * 36

LQ Regulator with Command Input (Proportional Control Law) u = u C Cx( t) How do we define u C? 37 Non-Zero Steady-State Regulation with LQ Regulator Command input provides equivalent state and control values for the LQ regulator Control law with command input x *( t) B 1 y * u = u* C x t = B y *C x t = ( B + CB 1 )y *Cx t 38

LQ Regulator with Forward Gain Matrix x *( t) u = u* C x t = C F y *C B x t C F B + CB 1 C B C Disturbance affects the system, whether or not it is measured If measured, disturbance effect of can be countered by C D (analogous to C F ) 39 Next Time: Cost Functions and Controller Structures 4

Supplemental Material 41 M t Square-Root Solution for the Algebraic Riccati Equation Q F T P PF + PGR 1 G T P = Square root of P: P DD T ; D P Integrate D to steady state D T t f D ( t) = D T M LT ( t), D t f M LT ( t) + M UT ( t) = D 1 ( t)f T D( t) D T t F T D T t D 1 t u = R 1 G T T D SS D SS x t = C SS x t where d 11 d D = 11 d 11 d 11 d 11 d 11 = P t f t f QD T t + D T t and GR 1 G T D T t i < j 1 ( m ij ) LT ( t) = m ij, i = j i > j m ij 4

Matrix Inverse Identity OCE, eq..-57 to -67 B 11 B 1 B 1 B A 11 A 1 A 1 A I m+n = I n I m B 11 B 1 B 1 B A 11 A 1 A 1 A ( B 11 A 1 + B 1 A ) ( B 1 A 1 + B A ) = B A + B A 11 11 1 1 B 1 A 11 + B A 1 ( B 11 A 11 + B 1 A 1 ) = I n ( B 11 A 1 + B 1 A ) = = B 1 A 11 + B A 1 ( B 1 A 1 + B A ) = I m Solve for B, then B 1 and B 1, then B 11 43