TUTORIAL on LOGIC-BASED CONTROL Part I: SWITCHED CONTROL SYSTEMS Daniel Liberzon Coordinated Science Laboratory and Dept. of Electrical & Computer Eng., Univ. of Illinois at Urbana-Champaign MED ’02, Lisbon OUTLINE Switched Control Systems Questions, Break Stability of Switched Systems OUTLINE Switched Control Systems Questions, Break Stability of Switched Systems SWITCHED and HYBRID SYSTEMS Switched systems: continuous systems with Hybrid systems: interaction of continuous anddiscrete discreteswitchings dynamics emphasis on properties of continuous state Switching can be: • State-dependent or Time-dependent x f (x ) where x f p (x), p P is a family of systems and : [0, ) P is a switching signal • Autonomous or Controlled SWITCHING CONTROL Plant: Classical continuous feedback paradigm: u P u y y P C u But logical decisions are often necessary: y P C1 C2 The closed-loop system is hybrid logic REASONS for SWITCHING • Nature of the control problem • Sensor or actuator limitations • Large modeling uncertainty • Combinations of the above REASONS for SWITCHING • Nature of the control problem • Sensor or actuator limitations • Large modeling uncertainty • Combinations of the above PARKING PROBLEM x2 x1 w1 cos x2 w1 sin x1 w2 Nonholonomic constraint: wheels do not slip OBSTRUCTION to STABILIZATION ? Solution: move away first REASONS for SWITCHING • Nature of the control problem • Sensor or actuator limitations • Large modeling uncertainty • Combinations of the above OUTPUT FEEDBACK Example: harmonic oscillator y y u u y u 12 y switched system QUANTIZED FEEDBACK q(x) u x PLANT QUANTIZER q(x) CONTROLLER x sensitivity M values OBSTRUCTION to STABILIZATION Assume: M, fixed Asymptotic stabilization is impossible MOTIVATING EXAMPLES 1. Temperature sensor normal too low 2. Camera with zoom Tracking a golf ball 3. Coding and decoding too high VARYING the SENSITIVITY zoom in zoom out Why switch ? • More realistic • Easier to design and analyze • Robust to time delays LINEAR SYSTEMS x Ax Bu Assume: K s.t. x Ax BKx is GAS: V 0 Along solutions of x Ax BK ( x e) quantization error we have |e| | x| V 0 Then can achieve GAS for some 0 SWITCHING POLICY |e| | x | V 0 We have | e | n / 2 level sets of V | x| n / 2 . NONLINEAR SYSTEMS x f ( x, u ) Assume: k s.t. x f ( x, k ( x )) is GAS: V 0 Need: along solutions of x f ( x, k ( x e)) |e | (| x |) V 0 where quantization error is pos. def., increasing, and unbounded (this is input-to-state stability wrt measurement error) Then can achieve GAS EXTENSIONS and APPLICATIONS • Arbitrary quantization regions • Active probing for information • Output and input quantization • Relaxing the assumptions • Performance-based design • Application to visual servoing REASONS for SWITCHING • Nature of the control problem • Sensor or actuator limitations • Large modeling uncertainty • Combinations of the above MODEL UNCERTAINTY unmodeled dynamics 0 parametric uncertainty p* P Also, noise and disturbances Adaptive control (continuous tuning) vs. supervisory control (switching) SUPERVISORY CONTROL Supervisor candidate controllers Controller 1 Controller 2 ... Controller m u1 u2 u um ... switching signal Plant y STABILITY of SWITCHED SYSTEMS x f (x) unstable Stable if: • switching stops in finite time • slow switching (on the average) • “locally confined” switching • common Lyapunov function REASONS for SWITCHING • Nature of the control problem • Sensor or actuator limitations • Large modeling uncertainty • Combinations of the above PARKING PROBLEM under UNCERTAINTY x2 x1 p1 w1 cos x2 p1 w1 sin p2 w2 x1 Unknown parameters p1 , p 2 correspond to the radius of rear wheels and distance between them SIMULATION OUTLINE Switched Control Systems Questions, Break Stability of Switched Systems OUTLINE Switched Control Systems Questions, Break Stability of Switched Systems TWO BASIC PROBLEMS • Stability for arbitrary switching • Stability for constrained switching UNIFORM STABILITY x f (x ) x f p (x), p P where is a family of GAS systems : [0, ) P is a piecewise constant switching signal Want GUAS : | x(t ) | ( | x(0) | , t ) t , ( KL) w.r.t. GUES: | x(t ) | c | x(0) | e t t , COMMON LYAPUNOV FUNCTION x f (x) is GUAS if and only if V s.t. V f p ( x) W ( x) 0 x 0, p P x V x Example: f p ( x) p x, P (0,1] (0,1) Corollary: is GAS f p ( x) 0 f px(1 0 , )pf is not enough q P p, q P t ( s ) ds 1 0 x(t ) e x(0) 0 if L Usually we take P compact and fp continuous SWITCHED LINEAR SYSTEMS x A x LAS for every GUES common Lyapunov function but not necessarily quadratic COMMUTING STABLE MATRICES => GUES P {1, 2}, A1 A2 A2 A1 1 s1 2 t1 1 s2 2 … t t2 A2 tk A1s k … A2 t1 A1s1 x(0) x(t ) e e e e A2 (tk ... t1) A1 ( s k ... s1) e e x(0) 0 quadratic common Lyap fcn: A1T P1 P1 A1 I A2T P2 P2 A2 P1 LIE ALGEBRAS and STABILITY g1 g, Lie algebra: g {Ap , p P} L A Lie bracket: [ A1 , A2 ] A1 A2 A2 A1 g k 1 [ g , g k ] g k g is nilpotent if k s.t. g k 0 (k ) g(1) g , g(k 1) [ g (k ,) g(k ) ] g (k ) g is solvable if k s.t. g 0 SOLVABLE LIE ALGEBRA => GUES g is solvable Lie’s Theorem: triangular form 1 Ap 0 * n quadratic common Lyap fcn: x TD x , D diagonal SOLVABLE + COMPACT => GUES Levi decomposition: g rs radical (max solvable ideal) s is compact => GUES quadratic common Lyap fcn SOLVABLE + NONCOMPACT => GUES g r s , I g , s is not compact • a set of stable generators for g that gives GUES • a set of stable generators for g that leads to an unstable system Lie algebra doesn’t provide enough information NONLINEAR SYSTEMS • Commuting systems [ f p , f q ] 0 => GUAS • Linearization • ??? REMARKS on LIE-ALGEBRAIC CRITERIA • Checkable conditions • Independent of representation • In terms of the original data • Not robust to small perturbations SYSTEMS with SPECIAL STRUCTURE • Triangular systems Linear => GUES Nonlinear: need ISS conditions u • Feedback systems Passive: V uy , K p 0 - => GUAS Small gain: || || 1, || K p || 1 => GUES • 2D systems convex combs of A1, A2 , A11 , A21 stable quadratic common Lyap fcn K y MULTIPLE LYAPUNOV FUNCTIONS x f1 ( x) , x f 2 ( x) GAS V1 , V2 respective Lyapunov functions V (t ) (t ) x f (x) is GAS 1 2 1 2 t Very useful for analysis of state-dependent switching MULTIPLE LYAPUNOV FUNCTIONS decreasing sequence V (t ) (t ) x f (x) is GAS decreasing sequence 1 2 1 2 t DWELL TIME The switching times t1 , t 2 , ... satisfy ti 1 ti D x f1 ( x) , x f 2 ( x) GES dwell time V1 Va11,| Vx2| functions f1 ( x) 1V1 ( x) Vrespective , 1 ( x) b1 | x |Lyapunov x V2 2 2 a2 | x | V2 ( x) b2 | x | , f 2 ( x) 2V2 ( x) x 2 2 must be Need: V1 (t 2 ) V1 (t0 ) 1 b1 b1 2 D V1 (t 2 ) V2 (t 2 ) e V2 (t1 ) a2 a2 b1 b2 2 D b1 b2 (1 2 ) D e V1 (t1 ) e V1 (t 0 ) a2 a a a 2 1 1 1 t0 1 2 t1 t2 t AVERAGE DWELL TIME N (T , t ) N 0 T t AD # of switches on (t , T ) average dwell time N 0 0 no switching: cannot switch if T t AD N 0 1 dwell time: cannot switch twice if T t AD x f (x) 1 (| x | ) V p ( x) 2 (| x | ) Vp f ( x ) V p ( x ) x p x f (x) V p ( x) Vq ( x), p, q P => is GAS if AD log SWITCHED LINEAR SYSTEMS x A x • GUES over all with large enough AD • Finite induced norms for x A x B u y C x • The case when some subsystems are unstable STABILIZATION by SWITCHING x A1x , x A2 x both unstable Assume: A A1 (1 ) A2 stable for some (0,1) T (1 )( A T P PA ) 0 ( A1T P PA1A ) P PA 20 2 So for each x 0: either xT ( A1T P PA1 ) x 0 or xT ( A2T P PA2 ) x 0 UNSTABLE CONVEX COMBINATIONS Can also use multiple Lyapunov functions LMIs