Gradient Projection Anti-windup Scheme Thesis Defense Justin Teo (MIT Aero/Astro) Thesis Committee: Thesis Readers: Department Representative: Jonathan P. How (Chair) Emilio Frazzoli Steven R. Hall Eugene Lavretsky Luca F. Bertuccelli Louis Breger (MIT Aero/Astro) (MIT Aero/Astro) (MIT Aero/Astro) (Boeing) (MIT Aero/Astro) (Draper) Wesley L. Harris (MIT Aero/Astro) December 20, 2010 Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 1 / 35 Outline Outline 1 Introduction 2 GPAW Compensated Controller 3 Input Constrained Planar LTI Systems 4 An ROA Comparison Result 5 A Numerical Comparison 6 Conclusions Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 2 / 35 Introduction Effects of Control Saturation Effects of Control Saturation Well Recognized Fact [Bernstein and Michel 1995] Control saturation affects virtually all practical control systems Effects called “windup”, affects all dynamic controllers and leads to: performance degradation (with certainty) instability (possibly) Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 3 / 35 Introduction Effects of Control Saturation Effects of Control Saturation Well Recognized Fact [Bernstein and Michel 1995] Control saturation affects virtually all practical control systems Effects called “windup”, affects all dynamic controllers and leads to: performance degradation (with certainty) instability (possibly) Stable Plant, Unstable Controller Mild effects [Visioli 2006]: large overshoots long settling times unconstrained saturated 2 x sluggish response 4 0 −2 0 2 4 6 8 10 12 14 16 18 20 18 20 2 umax u 1 0 −1 umin −2 −3 0 2 4 6 8 10 12 14 16 time (s) Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 3 / 35 Introduction Effects of Control Saturation Effects of Control Saturation Well Recognized Fact [Bernstein and Michel 1995] Control saturation affects virtually all practical control systems Effects called “windup”, affects all dynamic controllers and leads to: performance degradation (with certainty) instability (possibly) Stable Plant, Unstable Controller Mild effects [Visioli 2006]: large overshoots long settling times unconstrained saturated 2 x sluggish response 4 0 −2 0 2 4 6 8 10 12 14 18 20 18 20 umax u 1 0 −1 Severe effects: instability 16 2 umin −2 −3 0 2 4 6 8 10 12 14 16 time (s) Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 3 / 35 Introduction Effects of Control Saturation Disasters Caused Indirectly by Windup Disasters caused indirectly by windup include: 1986 Chernobyl (nuclear reactor) disaster [Stein 2003] 1992 crash of YF-22 fighter aircraft [Dornheim 1992] 1989 and 1993 crashes of Saab Gripen JAS 39 fighter aircraft [Butterworth-Hayes 1994, Stein 2003] 1992 crash of YF-22 Justin Teo (ACL, MIT) 1989 and 1993 crashes of Saab Gripen Gradient Projection Anti-windup Scheme Dec. 20, 2010 4 / 35 Introduction Control Design Strategies Control Design Strategies Control design strategies to deal with windup: avoiding saturation - applies when control task is well-defined, e.g. assembly lines Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 5 / 35 Introduction Control Design Strategies Control Design Strategies Control design strategies to deal with windup: avoiding saturation - applies when control task is well-defined, e.g. assembly lines one-step approach accounts for saturation in design of nominal controller - complex often conservative and hard to tune [Tarbouriech and Turner 2009, Sofrony et al. 2006, Mulder et al. 2009] Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 5 / 35 Introduction Control Design Strategies Control Design Strategies Control design strategies to deal with windup: avoiding saturation - applies when control task is well-defined, e.g. assembly lines one-step approach accounts for saturation in design of nominal controller - complex often conservative and hard to tune [Tarbouriech and Turner 2009, Sofrony et al. 2006, Mulder et al. 2009] two-step approach or anti-windup compensation ignores saturation in design of nominal controller (step 1) design controller modifications to account for windup (step 2) Anti-windup compensation preferred by practitioners due to [Tarbouriech and Turner 2009]: design of nominal controller greatly simplified can be retrofitted to existing controllers Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 5 / 35 Introduction Control Design Strategies Anti-windup Compensation Anti-windup compensation well studied for linear time invariant (LTI) case [Kothare et al. 1994, Edwards and Postlethwaite 1998, Tarbouriech and Turner 2009] r Σ̃c ũ u sat(u) v ẋ = Ax + Bv y y = Cx + Dv Unconstrained plant Σ̃aw Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 6 / 35 Introduction Control Design Strategies Anti-windup Compensation Anti-windup compensation well studied for linear time invariant (LTI) case [Kothare et al. 1994, Edwards and Postlethwaite 1998, Tarbouriech and Turner 2009] r Σ̃c ũ u sat(u) v y = Cx + Dv − yaw2 yaw1 Σ̃aw y ẋ = Ax + Bv Unconstrained plant w Anti-windup compensated controller Anti-windup compensator driven by w = sat(u) − u Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 6 / 35 Introduction Control Design Strategies Anti-windup Compensation Anti-windup compensation well studied for linear time invariant (LTI) case [Kothare et al. 1994, Edwards and Postlethwaite 1998, Tarbouriech and Turner 2009] r Σ̃c ũ u sat(u) y = Cx + Dv − yaw2 yaw1 Σ̃aw y ẋ = Ax + Bv v Unconstrained plant w Anti-windup compensated controller Anti-windup compensator driven by w = sat(u) − u Open Problem [Tarbouriech and Turner 2009] Anti-windup compensation for saturated nonlinear systems Most practical control systems are nonlinear - LTI are approximations Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 6 / 35 Introduction Problem Statement Problem Statement Nominal system Σn : feedback interconnection (FI) of Σp , Σc Saturated Plant: ( ẋ = f (x, sat(u)) Σp : y = g(x, sat(u)) Anti-windup (AW) compensated system Σaws : FI of Σp , Σaw Nominal Controller: ( ẋc = fc (xc , y, r) Σc : u = gc (xc , y, r) General Anti-windup Problem Design Σaw and determined initialization xaw (0) such that Σaws satisfies: AW Compensated Controller: ( ẋaw = faw (xaw , y, r) Σaw : u = gaw (xaw , y, r) Justin Teo (ACL, MIT) when no controls saturate for Σn , then nominal performance recovered, i.e. Σaws ≡ Σn when some controls saturate, stability and performance of Σaws is no worse than that of Σn Gradient Projection Anti-windup Scheme Dec. 20, 2010 7 / 35 Introduction Literature Review Literature Review Anti-windup methods (partial citations) applicable to nonlinear systems: Conditioning Technique [Hanus et al. 1987] - computationally prohibitive and severely limited for nonlinear systems Feedback Linearizable Nonlinear Systems [Yoon et al. 2008] - requires feedback linearizable plant and feedback linearizing controller For some Particular Controllers [Hu and Rangaiah 2000, Johnson and Calise 2001, 2003, Do et al. 2004] - not general purpose Nonlinear Anti-windup for Euler-Lagrange Systems [Morabito et al. 2004] - hard to generalize Optimal Directionality Compensation [Soroush and Daoutidis 2002] plant needs to be square Reference Governor [Gilbert and Kolmanovsky 2002] - some conservatism introduced Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 8 / 35 Introduction Contributions Contributions Contributions of this research include: developed general purpose anti-windup scheme motivated new paradigm for anti-windup problem demonstrated need to consider asymmetric saturation constraints for general saturated systems developed region of attraction (ROA) comparison and stability results for GPAW compensated (nonlinear) systems demonstrated viability of GPAW scheme as a candidate anti-windup scheme for general systems related GPAW compensated systems to projected dynamical systems and linear systems with partial state constraints Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 9 / 35 Introduction Contributions Contributions Contributions of this research include: developed general purpose anti-windup scheme motivated new paradigm for anti-windup problem demonstrated need to consider asymmetric saturation constraints for general saturated systems developed region of attraction (ROA) comparison and stability results for GPAW compensated (nonlinear) systems demonstrated viability of GPAW scheme as a candidate anti-windup scheme for general systems related GPAW compensated systems to projected dynamical systems and linear systems with partial state constraints Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 9 / 35 GPAW Compensated Controller Outline 1 Introduction 2 GPAW Compensated Controller 3 Input Constrained Planar LTI Systems 4 An ROA Comparison Result 5 A Numerical Comparison 6 Conclusions Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 10 / 35 GPAW Compensated Controller Conditional Integration Conditional Integration Conditional integration (CI) for PID controllers [Fertik and Ross 1967] 0, if u ≥ umax ∧ e > 0 ėi = 0, if u ≤ umin ∧ e < 0 ėi = e CI −→ e, otherwise u = Kp e + Ki ei + Kd ė u = Kp e + Ki ei + Kd ė Stop integration when nominal update will aggravate saturation constraints, or stop integration when departing unsaturated region K(e, ė) = {ēi ∈ R | sat(Kp e + Ki ēi + Kd ė) = Kp e + Ki ēi + Kd ė} Attempts to achieve controller state-output consistency sat(u) = u Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 11 / 35 GPAW Compensated Controller Conditional Integration Conditional Integration Conditional integration (CI) for PID controllers [Fertik and Ross 1967] 0, if u ≥ umax ∧ e > 0 ėi = 0, if u ≤ umin ∧ e < 0 ėi = e CI −→ e, otherwise u = Kp e + Ki ei + Kd ė u = Kp e + Ki ei + Kd ė Stop integration when nominal update will aggravate saturation constraints, or stop integration when departing unsaturated region K(e, ė) = {ēi ∈ R | sat(Kp e + Ki ēi + Kd ė) = Kp e + Ki ēi + Kd ė} Attempts to achieve controller state-output consistency sat(u) = u ẋci = fci (xci , y, r) Extends easily to decoupled nonlinear controllers uci = gci (xci , y, r) ẋc = fc (xc , y, r) For coupled nonlinear controllers, need projection opuc = gc (xc , y, r) erator - project onto K(e, ė) analogue Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 11 / 35 GPAW Compensated Controller Gradient Projection Method for Nonlinear Programming Gradient Projection Method for Nonlinear Programming [Rosen 1960, 1961] −∇J(x2 ) Nonlinear program: minq J(x) −∇J(x1 ) x∈R h̃(x) ≤ 0 x3 −∇J(x0 ) z3 x0 Projections: z1 , z2 , z3 Justin Teo (ACL, MIT) −∇J(x3 ) K̃ Gradient Projection Anti-windup Scheme x 3) H 3( G3 Boundaries: H1 , H2 , G3 z2 (x 3 ) Feasible region: K̃ = {x̄ | h̃(x̄) ≤ 0} x1 z 1 H1 zd H2 ∇h̃2 ∇h̃ 3 subject to x2 ∇h̃ 1 Dec. 20, 2010 12 / 35 GPAW Compensated Controller Gradient Projection Method for Nonlinear Programming Gradient Projection Method for Nonlinear Programming [Rosen 1960, 1961] −∇J(x2 ) Nonlinear program: minq J(x) −∇J(x1 ) x∈R h̃(x) ≤ 0 x3 −∇J(x0 ) z3 −∇J(x3 ) K̃ x0 Projections: z1 , z2 , z3 x 3) H 3( G3 Boundaries: H1 , H2 , G3 z2 (x 3 ) Feasible region: K̃ = {x̄ | h̃(x̄) ≤ 0} x1 z 1 H1 zd H2 ∇h̃2 ∇h̃ 3 subject to x2 ∇h̃ 1 Extended to continuous-time to yield projection operator - requires solution to combinatorial optimization subproblem at each point Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 12 / 35 GPAW Compensated Controller GPAW Compensated Controller GPAW Compensated Controller Gradient projection anti-windup (GPAW) compensated controller: obtained by applying projection operator from continuous-time gradient projection method on nominal controller defined by online solution to a combinatorial optimization subproblem Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 13 / 35 GPAW Compensated Controller GPAW Compensated Controller GPAW Compensated Controller Gradient projection anti-windup (GPAW) compensated controller: obtained by applying projection operator from continuous-time gradient projection method on nominal controller defined by online solution to a combinatorial optimization subproblem For “strictly proper” nonlinear controllers, ẋc = fc (xc , y, r) uc = gc (xc ) GPAW,Γ=ΓT >0 −−−−−−−−−−→ ẋg = RI ∗ (xg , y, r)fc (xg , y, r) ug = gc (xg ) Everything rests on projection operator RI ∗ ! Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 13 / 35 GPAW Compensated Controller GPAW Compensated Controller GPAW Compensated Controller Gradient projection anti-windup (GPAW) compensated controller: obtained by applying projection operator from continuous-time gradient projection method on nominal controller defined by online solution to a combinatorial optimization subproblem For “strictly proper” nonlinear controllers, ẋc = fc (xc , y, r) uc = gc (xc ) GPAW,Γ=ΓT >0 −−−−−−−−−−→ ẋg = fc (xg , y, r) ug = gc (xg ) Everything rests on projection operator RI ∗ ! Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 13 / 35 GPAW Compensated Controller GPAW Compensated Controller GPAW Compensated Controller Gradient projection anti-windup (GPAW) compensated controller: obtained by applying projection operator from continuous-time gradient projection method on nominal controller defined by online solution to a combinatorial optimization subproblem For “strictly proper” nonlinear controllers, ẋc = fc (xc , y, r) uc = gc (xc ) GPAW,Γ=ΓT >0 −−−−−−−−−−→ ẋg = RI ∗ (xg , y, r)fc (xg , y, r) ug = gc (xg ) Everything rests on projection operator RI ∗ ! Projection operator RI ∗ defined by Γ = ΓT > 0, online solution to combinatorial optimization subproblem I ∗ , and projection matrix ( I − ΓNI (NIT ΓNI )−1 NIT (xg ), if I = 6 ∅ RI (xg ) = I, otherwise Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 13 / 35 GPAW Compensated Controller GPAW Compensated Controller Other Properties GPAW compensated controller has a single parameter Γ = ΓT > 0: can be defined equivalently by online (unique) solution to: convex quadratic program (with numerous efficient solvers) projection onto convex polyhedral cone (algorithms available) - valid regardless of nonlinearities in plant/controller Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 14 / 35 GPAW Compensated Controller GPAW Compensated Controller Other Properties GPAW compensated controller has a single parameter Γ = ΓT > 0: can be defined equivalently by online (unique) solution to: convex quadratic program (with numerous efficient solvers) projection onto convex polyhedral cone (algorithms available) - valid regardless of nonlinearities in plant/controller can be realized by closed-form expressions when uc ∈ R or uc ∈ R2 - computationally efficient Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 14 / 35 GPAW Compensated Controller GPAW Compensated Controller Other Properties GPAW compensated controller has a single parameter Γ = ΓT > 0: can be defined equivalently by online (unique) solution to: convex quadratic program (with numerous efficient solvers) projection onto convex polyhedral cone (algorithms available) - valid regardless of nonlinearities in plant/controller can be realized by closed-form expressions when uc ∈ R or uc ∈ R2 - computationally efficient attempts to enforce control saturation constraints h i g (x )−u h(xg ) = −gcc (xgg )+umax ≤0 ⇔ xg ∈ K := {x̄ | h(x̄) ≤ 0} min Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 14 / 35 GPAW Compensated Controller GPAW Compensated Controller Other Properties GPAW compensated controller has a single parameter Γ = ΓT > 0: can be defined equivalently by online (unique) solution to: convex quadratic program (with numerous efficient solvers) projection onto convex polyhedral cone (algorithms available) - valid regardless of nonlinearities in plant/controller can be realized by closed-form expressions when uc ∈ R or uc ∈ R2 - computationally efficient attempts to enforce control saturation constraints h i g (x )−u h(xg ) = −gcc (xgg )+umax ≤0 ⇔ xg ∈ K := {x̄ | h(x̄) ≤ 0} min Non-“strictly proper” nonlinear controllers can be approximated arbitrarily well to be “strictly proper” (singular perturbation theory [Khalil 2002]) ẋc = fc (xc , y, r) uc = gc (xc , y, r) Justin Teo (ACL, MIT) a∈(0,∞) −−−−−→ ≈ x̃˙ c = f˜c (x̃c , y, r) uc = g̃c (x̃c ) Gradient Projection Anti-windup Scheme Dec. 20, 2010 14 / 35 GPAW Compensated Controller Controller State-output Consistency Controller State-output Consistency Controller State-output Consistency sat(gc (xg )) ≡ gc (xg ) ⇔ sat(ug ) ≡ ug ⇔ xg (t) ∈ K, ∀t ∈ R - implicit objective of anti-windup schemes (majority) driven by signal Figure (sat(u) − u) Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 15 / 35 GPAW Compensated Controller Controller State-output Consistency Controller State-output Consistency Controller State-output Consistency sat(gc (xg )) ≡ gc (xg ) ⇔ sat(ug ) ≡ ug ⇔ xg (t) ∈ K, ∀t ∈ R - implicit objective of anti-windup schemes (majority) driven by signal Figure (sat(u) − u) Theorem (GPAW Controller State-output Consistency) For GPAW compensated controller, if there exists a T ∈ R such that sat(ug (T )) = ug (T ), then sat(ug (t)) = ug (t) holds for all t ≥ T Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 15 / 35 GPAW Compensated Controller Controller State-output Consistency Controller State-output Consistency Controller State-output Consistency sat(gc (xg )) ≡ gc (xg ) ⇔ sat(ug ) ≡ ug ⇔ xg (t) ∈ K, ∀t ∈ R - implicit objective of anti-windup schemes (majority) driven by signal Figure (sat(u) − u) Theorem (GPAW Controller State-output Consistency) For GPAW compensated controller, if there exists a T ∈ R such that sat(ug (T )) = ug (T ), then sat(ug (t)) = ug (t) holds for all t ≥ T Implications - GPAW compensated closed-loop system: ẋ = f (x, sat(gc (xg ))) ẋg = RI ∗ fc (x, xg , sat(gc (xg )), r) Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 15 / 35 GPAW Compensated Controller Controller State-output Consistency Controller State-output Consistency Controller State-output Consistency sat(gc (xg )) ≡ gc (xg ) ⇔ sat(ug ) ≡ ug ⇔ xg (t) ∈ K, ∀t ∈ R - implicit objective of anti-windup schemes (majority) driven by signal Figure (sat(u) − u) Theorem (GPAW Controller State-output Consistency) For GPAW compensated controller, if there exists a T ∈ R such that sat(ug (T )) = ug (T ), then sat(ug (t)) = ug (t) holds for all t ≥ T Implications - GPAW compensated closed-loop system: ẋ = f (x, sat(gc (xg ))) xg (0)∈K ẋg = RI ∗ fc (x, xg , sat(gc (xg )), r) −−−−−→ ẋ = f (x, gc (xg )) ẋg = RI ∗ fc (x, xg , gc (xg ), r) saturation function sat(·) eliminated: significant simplification all complications arising from saturation accounted for by RI ∗ Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 15 / 35 GPAW Compensated Controller Controller State-output Consistency GPAW Scheme Visualization Nominal controller: ẋc = fc (xc , y, r) fc2 fc1 uc = gc (xc ) H2 ∇h2 xg3 fc0 fc3 H 3( K f g3 Boundaries: H1 , H2 , G3 xg1 f g1 fg2 ) H1 xg2 (x g3 ug = gc (xg ) ∇h 1 ∇h 3 GPAW controller: ẋg = RI ∗ fc (xg , y, r) xg0 ) Justin Teo (ACL, MIT) x g3 Unsaturated region: Nominal update: Projections: G3 Gradients: ∇hi (xg ) = ±∇gci (xg ) K := {x̄ | sat(gc (x̄)) = gc (x̄)} fci := fc (xgi , y(ti ), r(ti )) for xgi := xg (ti ) fgi := RI ∗ fci Gradient Projection Anti-windup Scheme Dec. 20, 2010 16 / 35 Input Constrained Planar LTI Systems Outline 1 Introduction 2 GPAW Compensated Controller 3 Input Constrained Planar LTI Systems 4 An ROA Comparison Result 5 A Numerical Comparison 6 Conclusions Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 17 / 35 Input Constrained Planar LTI Systems Projected Dynamical System Input Constrained Planar LTI Systems Simplest possible feedback system, 1st order LTI plant and controller: Σplant : ẋ = ax + b sat(u), GPAW compensated 0, u̇ = 0, cx + du, Justin Teo (ACL, MIT) u̇ = cx + du system: if u ≥ umax , cx + du > 0 if u ≤ umin , cx + du < 0 otherwise Gradient Projection Anti-windup Scheme Dec. 20, 2010 18 / 35 Input Constrained Planar LTI Systems Projected Dynamical System Input Constrained Planar LTI Systems Simplest possible feedback system, 1st order LTI plant and controller: Σplant : ẋ = ax + b sat(u), GPAW compensated 0, u̇ = 0, cx + du, u̇ = cx + du feedback −−−−−→ Σn feedback Σg Σplant system: if u ≥ umax , cx + du > 0 if u ≤ umin , cx + du < 0 otherwise −−−−−→ Σplant Assumption (Unconstrained Stability) The unconstrained system Σu (umax = −umin = ∞) is globally stable Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 18 / 35 Input Constrained Planar LTI Systems Projected Dynamical System Input Constrained Planar LTI Systems Simplest possible feedback system, 1st order LTI plant and controller: Σplant : ẋ = ax + b sat(u), GPAW compensated 0, u̇ = 0, cx + du, u̇ = cx + du feedback −−−−−→ Σn feedback Σg Σplant system: if u ≥ umax , cx + du > 0 if u ≤ umin , cx + du < 0 otherwise −−−−−→ Σplant Assumption (Unconstrained Stability) The unconstrained system Σu (umax = −umin = ∞) is globally stable Proposition (Relation to Projected Dynamical Systems) The GPAW compensated system Σg is a projected dynamical system [Dupuis and Nagurney 1993, Zhang and Nagurney 1995] Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 18 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Region of Attraction Containment Region of Attraction (ROA) limits utility of systems, defined as: Rn := {z̄ ∈ R2 | lim φn (t, z̄) = 0} t→∞ Justin Teo (ACL, MIT) Rg := {z̄ ∈ R2 | lim φg (t, z̄) = 0} Gradient Projection Anti-windup Scheme t→∞ Dec. 20, 2010 19 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Region of Attraction Containment Region of Attraction (ROA) limits utility of systems, defined as: Rn := {z̄ ∈ R2 | lim φn (t, z̄) = 0} t→∞ Rg := {z̄ ∈ R2 | lim φg (t, z̄) = 0} t→∞ Anti-windup schemes aim to improve performance only when saturated Require ROA to be maintained/enlarged to be valid anti-windup scheme, i.e. Rn ⊂ Raw Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 19 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Region of Attraction Containment Region of Attraction (ROA) limits utility of systems, defined as: Rn := {z̄ ∈ R2 | lim φn (t, z̄) = 0} t→∞ Rg := {z̄ ∈ R2 | lim φg (t, z̄) = 0} t→∞ Anti-windup schemes aim to improve performance only when saturated Require ROA to be maintained/enlarged to be valid anti-windup scheme, i.e. Rn ⊂ Raw Proposition (ROA Containment) The ROA of the origin of system Σn is contained within the ROA of the origin of system Σg , i.e. Rn ⊂ Rg ROA containment is a strong result: valid for all system parameters and saturation limits independent of any Lyapunov function implies for every Lyapunov function Vn ⇒ Rn , then ∃Vg ⇒ Rg (⊃ Rn ) stark departure from existing stability results Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 19 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Numerical Results I, Rn = Rg Unstable plant, stable controller, umax = −umin = 1 5 4 3 2 u 1 0 −1 −2 −3 −4 −5 −2.5 Justin Teo (ACL, MIT) −2 −1.5 −1 −0.5 0 x 0.5 1 Gradient Projection Anti-windup Scheme 1.5 2 2.5 Dec. 20, 2010 20 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Numerical Results I, Rn = Rg Unstable plant, stable controller, umax = −umin = 1 5 4 3 2 1 u Rn 0 −1 −2 −3 −4 −5 −2.5 Justin Teo (ACL, MIT) −2 −1.5 −1 −0.5 0 x 0.5 1 Gradient Projection Anti-windup Scheme 1.5 2 2.5 Dec. 20, 2010 20 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Numerical Results I, Rn = Rg Unstable plant, stable controller, umax = −umin = 1 5 4 3 2 1 u Rg 0 −1 −2 −3 −4 −5 −2.5 Justin Teo (ACL, MIT) −2 −1.5 −1 −0.5 0 x 0.5 1 Gradient Projection Anti-windup Scheme 1.5 2 2.5 Dec. 20, 2010 20 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Numerical Results I, Rn = Rg Unstable plant, stable controller, umax = −umin = 1 5 4 3 2 u 1 Rn = Rg 0 −1 −2 −3 −4 −5 −2.5 Justin Teo (ACL, MIT) −2 −1.5 −1 −0.5 0 x 0.5 1 Gradient Projection Anti-windup Scheme 1.5 2 2.5 Dec. 20, 2010 20 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Numerical Results I, Rn = Rg Unstable plant, stable controller, umax = −umin = 1 5 φn (t, z0 ) 4 φg (t, z0 ) 3 2 u 1 Rn = Rg 0 −1 −2 −3 −4 −5 −2.5 Justin Teo (ACL, MIT) −2 −1.5 −1 −0.5 0 x 0.5 1 Gradient Projection Anti-windup Scheme 1.5 2 2.5 Dec. 20, 2010 20 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Numerical Results I, Rn = Rg Unstable plant, stable controller, umax = −umin = 1 5 φn (t, z0 ) 4 φg (t, z0 ) 3 2 u 1 Rn = Rg 0 −1 −2 −3 −4 −5 −2.5 Justin Teo (ACL, MIT) −2 −1.5 −1 −0.5 0 x 0.5 1 Gradient Projection Anti-windup Scheme 1.5 2 2.5 Dec. 20, 2010 20 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Numerical Results I, Rn = Rg Unstable plant, stable controller, umax = −umin = 1 5 φn (t, z0 ) 4 φg (t, z0 ) 3 2 u 1 Rn = Rg 0 −1 −2 −3 −4 −5 −2.5 Justin Teo (ACL, MIT) −2 −1.5 −1 −0.5 0 x 0.5 1 Gradient Projection Anti-windup Scheme 1.5 2 2.5 Dec. 20, 2010 20 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Numerical Results I, Rn = Rg Unstable plant, stable controller, umax = −umin = 1 5 φn (t, z0 ) 4 φg (t, z0 ) 3 2 u 1 Rn = Rg 0 −1 −2 −3 −4 −5 −2.5 Justin Teo (ACL, MIT) −2 −1.5 −1 −0.5 0 x 0.5 1 Gradient Projection Anti-windup Scheme 1.5 2 2.5 Dec. 20, 2010 20 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Numerical Results II, Rn ⊂ Rg Unstable plant, stable controller, 1.5 = umax > −umin = 1 5 4 3 2 u 1 0 −1 −2 −3 −4 −5 −2.5 Justin Teo (ACL, MIT) −2 −1.5 −1 −0.5 0 x 0.5 1 Gradient Projection Anti-windup Scheme 1.5 2 2.5 Dec. 20, 2010 21 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Numerical Results II, Rn ⊂ Rg Unstable plant, stable controller, 1.5 = umax > −umin = 1 5 4 3 2 u 1 Rn 0 −1 −2 −3 −4 −5 −2.5 Justin Teo (ACL, MIT) −2 −1.5 −1 −0.5 0 x 0.5 1 Gradient Projection Anti-windup Scheme 1.5 2 2.5 Dec. 20, 2010 21 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Numerical Results II, Rn ⊂ Rg Unstable plant, stable controller, 1.5 = umax > −umin = 1 5 4 3 2 u 1 Rg 0 −1 −2 −3 −4 −5 −2.5 Justin Teo (ACL, MIT) −2 −1.5 −1 −0.5 0 x 0.5 1 Gradient Projection Anti-windup Scheme 1.5 2 2.5 Dec. 20, 2010 21 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Numerical Results II, Rn ⊂ Rg Unstable plant, stable controller, 1.5 = umax > −umin = 1 5 4 3 2 u 1 Rg 0 −1 −2 Rn −3 −4 −5 −2.5 Justin Teo (ACL, MIT) −2 −1.5 −1 −0.5 0 x 0.5 1 Gradient Projection Anti-windup Scheme 1.5 2 2.5 Dec. 20, 2010 21 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Numerical Results II, Rn ⊂ Rg Unstable plant, stable controller, 1.5 = umax > −umin = 1 5 φn (t, z0 ) 4 φg (t, z0 ) 3 2 u 1 Rg 0 −1 −2 Rn −3 −4 −5 −2.5 Justin Teo (ACL, MIT) −2 −1.5 −1 −0.5 0 x 0.5 1 Gradient Projection Anti-windup Scheme 1.5 2 2.5 Dec. 20, 2010 21 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Numerical Results II, Rn ⊂ Rg Unstable plant, stable controller, 1.5 = umax > −umin = 1 5 φn (t, z0 ) 4 φg (t, z0 ) 3 2 u 1 Rg 0 −1 −2 Rn −3 −4 −5 −2.5 Justin Teo (ACL, MIT) −2 −1.5 −1 −0.5 0 x 0.5 1 Gradient Projection Anti-windup Scheme 1.5 2 2.5 Dec. 20, 2010 21 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Numerical Results II, Rn ⊂ Rg Unstable plant, stable controller, 1.5 = umax > −umin = 1 5 φn (t, z0 ) 4 φg (t, z0 ) 3 2 u 1 Rg 0 −1 −2 Rn −3 −4 −5 −2.5 Justin Teo (ACL, MIT) −2 −1.5 −1 −0.5 0 x 0.5 1 Gradient Projection Anti-windup Scheme 1.5 2 2.5 Dec. 20, 2010 21 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Numerical Results II, Rn ⊂ Rg Unstable plant, stable controller, 1.5 = umax > −umin = 1 5 φn (t, z0 ) 4 φg (t, z0 ) 3 2 u 1 Rg 0 −1 −2 Rn −3 −4 −5 −2.5 Justin Teo (ACL, MIT) −2 −1.5 −1 −0.5 0 x 0.5 1 Gradient Projection Anti-windup Scheme 1.5 2 2.5 Dec. 20, 2010 21 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Numerical Results III, Rn ⊂ Rg Stable plant, unstable controller, umax = −umin 3 2 u 1 0 −1 −2 −3 −5 Justin Teo (ACL, MIT) −4 −3 −2 −1 0 x 1 2 Gradient Projection Anti-windup Scheme 3 4 5 Dec. 20, 2010 22 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Numerical Results III, Rn ⊂ Rg Stable plant, unstable controller, umax = −umin 3 Rn 2 u 1 0 −1 −2 −3 −5 Justin Teo (ACL, MIT) −4 −3 −2 −1 0 x 1 2 Gradient Projection Anti-windup Scheme 3 4 5 Dec. 20, 2010 22 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Numerical Results III, Rn ⊂ Rg Stable plant, unstable controller, umax = −umin 3 2 Rg u 1 0 −1 −2 −3 −5 Justin Teo (ACL, MIT) −4 −3 −2 −1 0 x 1 2 Gradient Projection Anti-windup Scheme 3 4 5 Dec. 20, 2010 22 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Numerical Results III, Rn ⊂ Rg Stable plant, unstable controller, umax = −umin 3 Rn 2 Rg u 1 0 −1 −2 −3 −5 Justin Teo (ACL, MIT) −4 −3 −2 −1 0 x 1 2 Gradient Projection Anti-windup Scheme 3 4 5 Dec. 20, 2010 22 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Numerical Results III, Rn ⊂ Rg Stable plant, unstable controller, umax = −umin 3 Rn 2 Rg u 1 0 −1 φn (t, z0 ) −2 φg (t, z0 ) −3 −5 Justin Teo (ACL, MIT) −4 −3 −2 −1 0 x 1 2 Gradient Projection Anti-windup Scheme 3 4 5 Dec. 20, 2010 22 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Numerical Results III, Rn ⊂ Rg Stable plant, unstable controller, umax = −umin 3 Rn 2 Rg u 1 0 −1 φn (t, z0 ) −2 φg (t, z0 ) −3 −5 Justin Teo (ACL, MIT) −4 −3 −2 −1 0 x 1 2 Gradient Projection Anti-windup Scheme 3 4 5 Dec. 20, 2010 22 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Numerical Results III, Rn ⊂ Rg Stable plant, unstable controller, umax = −umin 3 Rn 2 Rg u 1 0 −1 φn (t, z0 ) −2 φg (t, z0 ) −3 −5 Justin Teo (ACL, MIT) −4 −3 −2 −1 0 x 1 2 Gradient Projection Anti-windup Scheme 3 4 5 Dec. 20, 2010 22 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Numerical Results III, Rn ⊂ Rg Stable plant, unstable controller, umax = −umin 3 Rn 2 Rg u 1 0 −1 φn (t, z0 ) −2 φg (t, z0 ) −3 −5 Justin Teo (ACL, MIT) −4 −3 −2 −1 0 x 1 2 Gradient Projection Anti-windup Scheme 3 4 5 Dec. 20, 2010 22 / 35 Input Constrained Planar LTI Systems Region of Attraction Containment Numerical Results IV, Rn ⊂ Rg Unstable plant, stable controller Symmetric constraints Asymmetric constraints 5 5 φn (t, z0 ) 4 φn (t, z0 ) 4 φg (t, z0 ) φg (t, z0 ) 3 3 2 2 1 Rn = Rg 0 u u 1 Rg 0 −1 −1 −2 −2 −3 −3 −4 −4 −5 −2.5 −5 −2.5 −2 −1.5 −1 −0.5 0 x 0.5 1 1.5 2 2.5 Rn −2 −1.5 −1 −0.5 0 x 0.5 1 1.5 2 2.5 Stable plant, unstable controller 3 Rn 2 Rg u 1 0 −1 φn (t, z0 ) −2 φg (t, z0 ) −3 −5 Justin Teo (ACL, MIT) −4 −3 −2 −1 0 x 1 2 3 4 5 Gradient Projection Anti-windup Scheme Dec. 20, 2010 23 / 35 Input Constrained Planar LTI Systems New Paradigm for Anti-windup Problem New Paradigm for Anti-windup Problem Claim (Global Asymptotic Stability of Nominal System) If both open-loop plant and nominal controller are marginally or strictly stable, then the origin of Σn is globally asymptotically stable (GAS) and locally exponentially stable (LES), i.e. Rn = R2 Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 24 / 35 Input Constrained Planar LTI Systems New Paradigm for Anti-windup Problem New Paradigm for Anti-windup Problem Claim (Global Asymptotic Stability of Nominal System) If both open-loop plant and nominal controller are marginally or strictly stable, then the origin of Σn is globally asymptotically stable (GAS) and locally exponentially stable (LES), i.e. Rn = R2 Corollary (Global Asymptotic Stability of GPAW System) If both open-loop plant and nominal controller are marginally or strictly stable, then the origin of Σg is GAS and LES (Rg ⊃ Rn = R2 ) Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 24 / 35 Input Constrained Planar LTI Systems New Paradigm for Anti-windup Problem New Paradigm for Anti-windup Problem Claim (Global Asymptotic Stability of Nominal System) If both open-loop plant and nominal controller are marginally or strictly stable, then the origin of Σn is globally asymptotically stable (GAS) and locally exponentially stable (LES), i.e. Rn = R2 Corollary (Global Asymptotic Stability of GPAW System) If both open-loop plant and nominal controller are marginally or strictly stable, then the origin of Σg is GAS and LES (Rg ⊃ Rn = R2 ) Some anti-windup results are of the form of preceding Corollary Such results tells nothing about advantages of anti-windup method Same result obtained as Corollary of ROA containment ROA containment result shows true advantage Propose new paradigm to search for relative results Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 24 / 35 Input Constrained Planar LTI Systems Need to Consider Asymmetric Saturation Constraints Need to Consider Asymmetric Constraints Conjecture (Relaxing Constraints Imply ROA Enlargement) Let Rn1 be ROA for some saturation limits umin1 , umax1 , and Rn2 be ROA for umin2 , umax2 . If [umin1 , umax1 ] ⊂ [umin2 , umax2 ], then Rn1 ⊂ Rn2 Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 25 / 35 Input Constrained Planar LTI Systems Need to Consider Asymmetric Saturation Constraints Need to Consider Asymmetric Constraints Conjecture (Relaxing Constraints Imply ROA Enlargement) Let Rn1 be ROA for some saturation limits umin1 , umax1 , and Rn2 be ROA for umin2 , umax2 . If [umin1 , umax1 ] ⊂ [umin2 , umax2 ], then Rn1 ⊂ Rn2 Conjecture intuitively appealing, but WRONG! Symmetric Asymmetric 5 5 φn (t, z0 ) 4 φn (t, z0 ) 4 φg (t, z0 ) φg (t, z0 ) 3 3 2 2 1 Rn = Rg 0 u u 1 −1 −2 −2 −3 −3 −4 −4 −5 −2.5 −2 −1.5 −1 −0.5 0 x 0.5 1 1.5 umax = −umin = 1 2 2.5 −5 −2.5 Not pathological Rg 0 −1 Rn −2 −1.5 −1 −0.5 0 x 0.5 1 1.5 2 2.5 umax = 1.5, umin = −1 Need to consider asymmetric saturation constraints Most literature (less [Hu et al. 2002]) considers only symmetric constraints Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 25 / 35 An ROA Comparison Result Outline 1 Introduction 2 GPAW Compensated Controller 3 Input Constrained Planar LTI Systems 4 An ROA Comparison Result 5 A Numerical Comparison 6 Conclusions Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 26 / 35 An ROA Comparison Result An ROA Comparison Result Plant, Σp ẋ = f (x, sat(u)) Nominal Controller, Σc ẋc = fc (xc , y) y = g(x, sat(u)) u = gc (xc ) Justin Teo (ACL, MIT) GPAW Controller, Σgpaw ẋg = RI ∗ fc (xg , y) Gradient Projection Anti-windup Scheme u = gc (xg ) Dec. 20, 2010 27 / 35 An ROA Comparison Result An ROA Comparison Result Plant, Σp ẋ = f (x, sat(u)) Nominal Controller, Σc ẋc = fc (xc , y) GPAW Controller, Σgpaw ẋg = RI ∗ fc (xg , y) y = g(x, sat(u)) u = gc (xc ) u = gc (xg ) Assume zeq ∈ K \ ∂K is asymptotically stable equilibrium for Σn and Σg Nominal System: Σn : Σ p + Σ c ROA: Rn (zeq ) GPAW System: Σg : Σp + Σgpaw ROA: Rg (zeq ) Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 27 / 35 An ROA Comparison Result An ROA Comparison Result Plant, Σp ẋ = f (x, sat(u)) Nominal Controller, Σc ẋc = fc (xc , y) GPAW Controller, Σgpaw ẋg = RI ∗ fc (xg , y) y = g(x, sat(u)) u = gc (xc ) u = gc (xg ) Assume zeq ∈ K \ ∂K is asymptotically stable equilibrium for Σn and Σg Nominal System: Σn : Σ p + Σ c ROA: Rn (zeq ) GPAW System: Σg : Σp + Σgpaw ROA: Rg (zeq ) ROA estimate: ΩV = {z̄ | V (z̄) ≤ c} ⊂ Rn (zeq ) for some Lyapunov function V (z) = V (x, xc ) of Σn Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 27 / 35 An ROA Comparison Result An ROA Comparison Result Plant, Σp ẋ = f (x, sat(u)) Nominal Controller, Σc ẋc = fc (xc , y) GPAW Controller, Σgpaw ẋg = RI ∗ fc (xg , y) y = g(x, sat(u)) u = gc (xc ) u = gc (xg ) Assume zeq ∈ K \ ∂K is asymptotically stable equilibrium for Σn and Σg Nominal System: Σn : Σ p + Σ c ROA: Rn (zeq ) GPAW System: Σg : Σp + Σgpaw ROA: Rg (zeq ) ROA estimate: ΩV = {z̄ | V (z̄) ≤ c} ⊂ Rn (zeq ) for some Lyapunov function V (z) = V (x, xc ) of Σn Theorem (ROA Bounds for GPAW Compensated System) If there exists a Γ = ΓT > 0 such that ∂V (x̄, x̄c ) ∂V (x̄, x̄c ) RI ∗ fc ≤ fc , ∂xc ∂xc ∀(x̄, x̄c ) ∈ ΩV ∩ (Rn × K) then Σg with Γ has ROA satisfying (ΩV ∩ (Rn × K)) ⊂ Rg (zeq ) Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 27 / 35 An ROA Comparison Result ROA Comparison Indicates True Advantage Existing anti-windup results are in “absolute” sense may not indicate any advantages of anti-windup scheme ROA comparison result is in “relative” sense directly shows advantage of GPAW scheme first in new anti-windup paradigm Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 28 / 35 An ROA Comparison Result ROA Comparison Indicates True Advantage Existing anti-windup results are in “absolute” sense may not indicate any advantages of anti-windup scheme ROA comparison result is in “relative” sense directly shows advantage of GPAW scheme first in new anti-windup paradigm States loosely that ROA of Σg is not less than ROA estimate ΩV Applies for asymmetric saturation constraints Specialized with additional assumptions (e.g. LTI) Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 28 / 35 An ROA Comparison Result ROA Comparison Indicates True Advantage Existing anti-windup results are in “absolute” sense may not indicate any advantages of anti-windup scheme ROA comparison result is in “relative” sense directly shows advantage of GPAW scheme first in new anti-windup paradigm States loosely that ROA of Σg is not less than ROA estimate ΩV Applies for asymmetric saturation constraints Specialized with additional assumptions (e.g. LTI) Main condition: ∂V (x̄,x̄c ) ∗ ∂xc RI fc ≤ ∂V (x̄,x̄c ) ∂xc fc independent of sat(·) Can be used in two ways: comparison against ROA estimate of unconstrained system or nominal system Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 28 / 35 An ROA Comparison Result Application of ROA Comparison Result Example nonlinear planar system [Khalil 2002] ( ẋ = − sat(u) ( ẋ = − sat(u) Σn : Σ : 0, if A1 gs u̇ = x + (x2 − 1)u u̇ = 2 x + (x − 1)u, otherwise Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 29 / 35 An ROA Comparison Result Application of ROA Comparison Result Example nonlinear planar system [Khalil 2002] ( ẋ = −u ( ẋ = −u Σu : Σ : 0, if A1 g u̇ = x + (x2 − 1)u u̇ = 2 x + (x − 1)u, otherwise Compare with ROA estimate ΩV of unconstrained system: 4 3 uncompensated GPAW 2 x Ru (zeq ) 2 0 1 u ΩV −2 0 2 0 1 2 3 1 2 3 4 5 6 7 8 9 10 4 5 6 7 8 9 10 1 u −1 −2 −3 −3 Rg (zeq ) −2 uncompensated GPAW −1 Justin Teo (ACL, MIT) 0 x 1 2 0 −1 −2 0 3 Gradient Projection Anti-windup Scheme time (s) Dec. 20, 2010 29 / 35 An ROA Comparison Result Application of ROA Comparison Result Example nonlinear planar system [Khalil 2002] ( ẋ = −u ( ẋ = −u Σu : Σ : 0, if A1 g u̇ = x + (x2 − 1)u u̇ = 2 x + (x − 1)u, otherwise Compare with ROA estimate ΩV of unconstrained system: 4 3 uncompensated GPAW 2 x Ru (zeq ) 2 0 1 u ΩV −2 0 2 0 1 2 3 1 2 3 4 5 6 7 8 9 10 4 5 6 7 8 9 10 1 u −1 −2 −3 −3 Rg (zeq ) −2 uncompensated GPAW −1 0 x 1 2 0 −1 −2 0 3 time (s) Toy example defeats methods for LTI systems, feedback linearizable systems, and nonlinear anti-windup [Morabito et al. 2004] Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 29 / 35 A Numerical Comparison Outline 1 Introduction 2 GPAW Compensated Controller 3 Input Constrained Planar LTI Systems 4 An ROA Comparison Result 5 A Numerical Comparison 6 Conclusions Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 30 / 35 A Numerical Comparison Numerical Comparison with Robot Example ROA comparison and stability results still too conservative Compare GPAW vs [Yoon et al. 2008] (feedback linearizable systems) vs [Morabito et al. 2004] (nonlinear anti-windup) without stability guarantees Feedback linearizable nonlinear plant with disturbance input w [Yoon et al. 2008]: ( # " x2 ẋ1 Σp : ẋ = = −10x1 −0.1x31 −48.54x2 −w+sat(u) , y=x ẋ2 6.67(1+0.1 sin x ) 1 Feedback linearizing PID controller: Σc Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 31 / 35 A Numerical Comparison Numerical Comparison with Robot Example ROA comparison and stability results still too conservative Compare GPAW vs [Yoon et al. 2008] (feedback linearizable systems) vs [Morabito et al. 2004] (nonlinear anti-windup) without stability guarantees Feedback linearizable nonlinear plant with disturbance input w [Yoon et al. 2008]: ( # " x2 ẋ1 Σp : ẋ = = −10x1 −0.1x31 −48.54x2 −w+sat(u) , y=x ẋ2 6.67(1+0.1 sin x ) 1 Feedback linearizing PID controller: Σc Nominal system: Feedback linearized AW System [Yoon et al. 2008]: Nonlinear AW system [Morabito et al. 2004]: GPAW compensated system: Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme NS: Σp + Σc FL NAW GPAW Dec. 20, 2010 31 / 35 A Numerical Comparison disturbance GPAW Achieves Comparable Performance 200 100 0 −100 −200 0 2 4 6 2 4 6 8 10 12 14 8 10 12 14 output 1.5 NS FL 16 NAW 18 GPAW 20 1 0.5 0 0 16 18 20 time (s) GPAW achieves comparable performance with state-of-the-art methods Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 32 / 35 Conclusions Outline 1 Introduction 2 GPAW Compensated Controller 3 Input Constrained Planar LTI Systems 4 An ROA Comparison Result 5 A Numerical Comparison 6 Conclusions Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 33 / 35 Conclusions GPAW in Context In context (some dates from [Tarbouriech and Turner 2009]): problem as old as control theory itself (James Watt’s governor - 1788) windup problem recognized (1930s) ad-hoc schemes devised and adopted (LTI) (1930s) academic studies (1950s) provably stable “modern” anti-windup schemes (LTI) (late 1990s) Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 34 / 35 Conclusions GPAW in Context In context (some dates from [Tarbouriech and Turner 2009]): problem as old as control theory itself (James Watt’s governor - 1788) windup problem recognized (1930s) ad-hoc schemes devised and adopted (LTI) (1930s) academic studies (1950s) provably stable “modern” anti-windup schemes (LTI) (late 1990s) provably stable classes of nonlinear systems (mid 2000s) provably stable general nonlinear systems (GPAW - 2010) Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 34 / 35 Conclusions GPAW in Context In context (some dates from [Tarbouriech and Turner 2009]): problem as old as control theory itself (James Watt’s governor - 1788) windup problem recognized (1930s) ad-hoc schemes devised and adopted (LTI) (1930s) academic studies (1950s) provably stable “modern” anti-windup schemes (LTI) (late 1990s) provably stable classes of nonlinear systems (mid 2000s) provably stable general nonlinear systems (GPAW - 2010) less conservative stability results (???) Future work (partial list): search for less conservative stability results consider robustness issues due to presence of noise, disturbances, time delays, and unmodeled dynamics Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 34 / 35 Conclusions Conclusions Contributions of this research include: developed general purpose anti-windup scheme Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 35 / 35 Conclusions Conclusions Contributions of this research include: developed general purpose anti-windup scheme motivated new paradigm for anti-windup problem Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 35 / 35 Conclusions Conclusions Contributions of this research include: developed general purpose anti-windup scheme motivated new paradigm for anti-windup problem demonstrated need to consider asymmetric saturation constraints for general saturated systems Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 35 / 35 Conclusions Conclusions Contributions of this research include: developed general purpose anti-windup scheme motivated new paradigm for anti-windup problem demonstrated need to consider asymmetric saturation constraints for general saturated systems developed region of attraction (ROA) comparison and stability results for GPAW compensated (nonlinear) systems Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 35 / 35 Conclusions Conclusions Contributions of this research include: developed general purpose anti-windup scheme motivated new paradigm for anti-windup problem demonstrated need to consider asymmetric saturation constraints for general saturated systems developed region of attraction (ROA) comparison and stability results for GPAW compensated (nonlinear) systems demonstrated viability of GPAW scheme as a candidate anti-windup scheme for general systems Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 35 / 35 Conclusions Conclusions Contributions of this research include: developed general purpose anti-windup scheme motivated new paradigm for anti-windup problem demonstrated need to consider asymmetric saturation constraints for general saturated systems developed region of attraction (ROA) comparison and stability results for GPAW compensated (nonlinear) systems demonstrated viability of GPAW scheme as a candidate anti-windup scheme for general systems related GPAW compensated systems to projected dynamical systems and linear systems with partial state constraints Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 35 / 35 Conclusions Conclusions Contributions of this research include: developed general purpose anti-windup scheme motivated new paradigm for anti-windup problem demonstrated need to consider asymmetric saturation constraints for general saturated systems developed region of attraction (ROA) comparison and stability results for GPAW compensated (nonlinear) systems demonstrated viability of GPAW scheme as a candidate anti-windup scheme for general systems related GPAW compensated systems to projected dynamical systems and linear systems with partial state constraints Questions? Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 35 / 35 Backup Slides Backup Slides Backup slides Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 36 / 35 Backup Slides Dissertation Overview Dissertation Overview Covered Chapter 1, Introduction. Dissertation on gradient projection anti-windup (GPAW) scheme. Remaining chapters: Chapter 2 Construction and Fundamental Properties Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 37 / 35 Backup Slides Dissertation Overview Dissertation Overview Covered Chapter 1, Introduction. Dissertation on gradient projection anti-windup (GPAW) scheme. Remaining chapters: Chapter 2 Construction and Fundamental Properties Chapter 3 Input Constrained Planar LTI Systems Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 37 / 35 Backup Slides Dissertation Overview Dissertation Overview Covered Chapter 1, Introduction. Dissertation on gradient projection anti-windup (GPAW) scheme. Remaining chapters: Chapter 2 Construction and Fundamental Properties Chapter 3 Input Constrained Planar LTI Systems Chapter 4 Geometric Properties and Region of Attraction Comparison Results Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 37 / 35 Backup Slides Dissertation Overview Dissertation Overview Covered Chapter 1, Introduction. Dissertation on gradient projection anti-windup (GPAW) scheme. Remaining chapters: Chapter 2 Construction and Fundamental Properties Chapter 3 Input Constrained Planar LTI Systems Chapter 4 Geometric Properties and Region of Attraction Comparison Results Chapter 5 Input Constrained MIMO LTI Systems Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 37 / 35 Backup Slides Dissertation Overview Dissertation Overview Covered Chapter 1, Introduction. Dissertation on gradient projection anti-windup (GPAW) scheme. Remaining chapters: Chapter 2 Construction and Fundamental Properties Chapter 3 Input Constrained Planar LTI Systems Chapter 4 Geometric Properties and Region of Attraction Comparison Results Chapter 5 Input Constrained MIMO LTI Systems Chapter 6 Numerical Comparisons Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 37 / 35 Backup Slides Dissertation Overview Dissertation Overview Covered Chapter 1, Introduction. Dissertation on gradient projection anti-windup (GPAW) scheme. Remaining chapters: Chapter 2 Construction and Fundamental Properties Chapter 3 Input Constrained Planar LTI Systems Chapter 4 Geometric Properties and Region of Attraction Comparison Results Chapter 5 Input Constrained MIMO LTI Systems Chapter 6 Numerical Comparisons Chapter 7 Conclusions and Future Work Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 37 / 35 Backup Slides Dissertation Overview Dissertation Overview Covered Chapter 1, Introduction. Dissertation on gradient projection anti-windup (GPAW) scheme. Remaining chapters: Chapter 2 Construction and Fundamental Properties Chapter 3 Input Constrained Planar LTI Systems Chapter 4 Geometric Properties and Region of Attraction Comparison Results Chapter 5 Input Constrained MIMO LTI Systems Chapter 6 Numerical Comparisons Chapter 7 Conclusions and Future Work Appendix A Closed Form Expressions for Single-output GPAW Compensated Controllers Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 37 / 35 Backup Slides Dissertation Overview Dissertation Overview Covered Chapter 1, Introduction. Dissertation on gradient projection anti-windup (GPAW) scheme. Remaining chapters: Chapter 2 Construction and Fundamental Properties Chapter 3 Input Constrained Planar LTI Systems Chapter 4 Geometric Properties and Region of Attraction Comparison Results Chapter 5 Input Constrained MIMO LTI Systems Chapter 6 Numerical Comparisons Chapter 7 Conclusions and Future Work Appendix A Closed Form Expressions for Single-output GPAW Compensated Controllers Appendix B Closed Form Expressions for GPAW Compensated Controllers with Output of Dimension Two Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 37 / 35 Backup Slides Dissertation Overview Dissertation Overview Covered Chapter 1, Introduction. Dissertation on gradient projection anti-windup (GPAW) scheme. Remaining chapters: Chapter 2 Construction and Fundamental Properties Chapter 3 Input Constrained Planar LTI Systems Chapter 4 Geometric Properties and Region of Attraction Comparison Results Chapter 5 Input Constrained MIMO LTI Systems Chapter 6 Numerical Comparisons Chapter 7 Conclusions and Future Work Appendix A Closed Form Expressions for Single-output GPAW Compensated Controllers Appendix B Closed Form Expressions for GPAW Compensated Controllers with Output of Dimension Two Appendix C Procedure to Apply GPAW Compensation Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 37 / 35 Backup Slides Application on Nonlinear Two-link Robot Application on Nonlinear Two-link Robot Two-link robot (plant): Σplant : H(xt )ẍt + C(xt , ẋt )ẋt = sat(u) x2 Adaptive sliding-mode (nominal) controller: â˙ = −ΘY T s feedback −−−−−→ Σn Σplant uc = Y â − KD s Approximate nominal controller: x̃˙ c = −ΘY T s ẋaug = a(z(y, r) − xaug ) uc = Ŷ (xaug )x̃c − KD ŝ(xaug ) x1 ( ≡ ẋc = fc (xc , y, r) uc = gc (xc ) GPAW compensated controller: ẋg = RI ∗ fc (xg , y, r) ug = gc (xg ) Justin Teo (ACL, MIT) feedback −−−−−→ Σplant Gradient Projection Anti-windup Scheme Σg Movies Dec. 20, 2010 38 / 35 Backup Slides Passivity Properties Passivity Properties Decompose Γ = ΦΦT , define: PI (xg ) := Φ−1 RI (xg )Φ SI (xg ) := I − PI (xg ) Passivity and L2 -gain of Projection Operators PI ∗ (xg , y, r) and SI ∗ (xg , y, r) are passive and with L2 -gain less than 1 r RI ∗ fc (xg , y, r) ṽ Φ−1 PI ∗ Φ w̃ ẋg = w̃ u u = gc (xg ) xg y ẋ = f (x, ũ) y = g(x, ũ) ũ sat(u) GPAW modifies uncompensated system with passive operator Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 39 / 35 Backup Slides Passivity Properties Passivity Properties Decompose Γ = ΦΦT , define: PI (xg ) := Φ−1 RI (xg )Φ SI (xg ) := I − PI (xg ) Passivity and L2 -gain of Projection Operators PI ∗ (xg , y, r) and SI ∗ (xg , y, r) are passive and with L2 -gain less than 1 r fc (xg , y, r) ṽ Φ−1 v PI ∗ w Φ w̃ ẋg = w̃ u u = gc (xg ) xg y ẋ = f (x, ũ) y = g(x, ũ) ũ sat(u) GPAW modifies uncompensated system with passive operator Can derive passivity and small-gain based stability results Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 39 / 35 Backup Slides Geometric Properties Geometric Bounding Condition Let K be unsaturated region, K = {x̄ | sat(gc (x̄)) = gc (x̄)} Let fc (x, y, r), fg (x, y, r) = RI ∗ fc (x, y, r) be the vector fields of nominal and GPAW compensated controllers fc1 fg1 x ker(K) fc2 f g2 xker K Let Γ = ΓT > 0 be the GPAW parameter Theorem (Geometric Bounding Condition) If unsaturated region K is a star domain, then for any x ∈ K and any xker ∈ ker(K), hΓ−1 (x − xker ), fg (x, y, r)i ≤ hΓ−1 (x − xker ), fc (x, y, r)i holds for all (y, r) and all Γ = ΓT > 0 Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 40 / 35 Backup Slides Geometric Properties Star Domains Examples and counterexamples of star domains in R2 : Star, ker(Xi ) 6= ∅ NOT Star, ker(Xi ) = ∅ X6 = Y1 ∪ Y2 ker(X3 ) X4 ker(X1 ) X1 ker(X2 ) ker(X4 ) Y1 Y2 X4 X7 X2 X5 Any convex set X is also a star domain with ker(X) = X For any non-convex star domain, ker(X) is a strict subset of X If X is a star domain, then Rn × X is also a star domain with kernel ker(Rn × X) = Rn × ker(X) Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 41 / 35 Backup Slides Geometric Properties Geometric Interpretation hΓ−1 (x − xker ), fg (x, y, r)i ≤ hΓ−1 (x − xker ), fc (x, y, r)i fc1 pg 2 x p c2 1 Justin Teo (ACL, MIT) pg GPAW controller: fg = RI ∗ fc p c1 Nominal controller: fc fg1 ker(K) fc2 xker fg2 Gradient Projection Anti-windup Scheme K Dec. 20, 2010 42 / 35 Backup Slides Geometric Properties GPAW in Context Standard anti-windup structure: r Σ̃c ũ u sat(u) v yaw1 Σ̃aw y y = Cx + Dv − yaw2 ẋ = Ax + Bv Unconstrained plant w Anti-windup compensated controller Virtually all anti-windup schemes are variants of above GPAW scheme has additional “built-in” features GPAW has single parameter, only for “fine tuning” GPAW alone comparable to three state-of-the-art methods GPAW has potential to be developed into truly general purpose anti-windup scheme with better stability guarantees Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 43 / 35 Backup Slides Geometric Properties Conclusions Anti-windup compensation for nonlinear systems is an open problem Developed GPAW scheme, a general purpose anti-windup scheme: achieves controller state-output consistency several ways to realize defined by passive operator has clear geometric properties Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 44 / 35 Backup Slides Geometric Properties Conclusions Anti-windup compensation for nonlinear systems is an open problem Developed GPAW scheme, a general purpose anti-windup scheme: achieves controller state-output consistency several ways to realize defined by passive operator has clear geometric properties Strong results for planar LTI systems: ROA containment result independent of any Lyapunov function shows qualitative weaknesses of existing results motivated new anti-windup paradigm to search for “relative” results shows need to consider asymmetric saturation constraints establish link to projected dynamical systems Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 44 / 35 Backup Slides Geometric Properties Conclusions Anti-windup compensation for nonlinear systems is an open problem Developed GPAW scheme, a general purpose anti-windup scheme: achieves controller state-output consistency several ways to realize defined by passive operator has clear geometric properties Strong results for planar LTI systems: ROA containment result independent of any Lyapunov function shows qualitative weaknesses of existing results motivated new anti-windup paradigm to search for “relative” results shows need to consider asymmetric saturation constraints establish link to projected dynamical systems Derived ROA comparison and stability results - first results to directly indicate advantages of anti-windup Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 44 / 35 Backup Slides Geometric Properties Conclusions Anti-windup compensation for nonlinear systems is an open problem Developed GPAW scheme, a general purpose anti-windup scheme: achieves controller state-output consistency several ways to realize defined by passive operator has clear geometric properties Strong results for planar LTI systems: ROA containment result independent of any Lyapunov function shows qualitative weaknesses of existing results motivated new anti-windup paradigm to search for “relative” results shows need to consider asymmetric saturation constraints establish link to projected dynamical systems Derived ROA comparison and stability results - first results to directly indicate advantages of anti-windup Even without stability proofs, ad-hoc methods can be used to design GPAW controller yielding comparable performance with state-of-the-art anti-windup methods Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 44 / 35 Backup Slides References References I D. S. Bernstein and A. N. Michel. A chronological bibliography on saturating actuators. Int. J. Robust Nonlinear Control, 5(5): 375 – 380, 1995. doi: 10.1002/rnc.4590050502. P. Butterworth-Hayes. Gripen crash raises canard fears. Aerosp. Am., 32(2):10 – 11, Feb. 1994. H. M. Do, T. Başar, and J. Y. Choi. An anti-windup design for single input adaptive control systems in strict feedback form. In Proc. American Control Conf., volume 3, pages 2551 – 2556, Boston, MA, June/July 2004. M. A. Dornheim. Report pinpoints factors leading to YF-22 crash. Aviat. Week Space Technol., 137(19):53 – 54, Nov. 1992. P. Dupuis and A. Nagurney. Dynamical systems and variational inequalities. Ann. Oper. Res., 44(1):7 – 42, Feb. 1993. doi: 10.1007/BF02073589. C. Edwards and I. Postlethwaite. Anti-windup and bumpless-transfer schemes. Automatica, 34(2):199 – 210, Feb. 1998. doi: 10.1016/S0005-1098(97)00165-9. H. A. Fertik and C. W. Ross. Direct digital control algorithm with anti-windup feature. ISA Trans., 6(4):317 – 328, 1967. E. Gilbert and I. Kolmanovsky. Nonlinear tracking control in the presence of state and control constraints: a generalized reference governor. Automatica, 38(12):2063 – 2073, Dec. 2002. doi: 10.1016/S0005-1098(02)00135-8. R. Hanus, M. Kinnaert, and J.-L. Henrotte. Conditioning technique, a general anti-windup and bumpless transfer method. Automatica, 23(6):729 – 739, Nov. 1987. doi: 10.1016/0005-1098(87)90029-X. Q. Hu and G. P. Rangaiah. Anti-windup schemes for uncertain nonlinear systems. IET Control Theory Appl., 147(3):321 – 329, May 2000. doi: 10.1049/ip-cta:20000136. T. Hu, A. N. Pitsillides, and Z. Lin. Null controllability and stabilization of linear systems subject to asymmetric actuator saturation. In V. Kapila and K. M. Grigoriadis, editors, Actuator Saturation Control, Control Eng., chapter 3, pages 47 – 76. Marcel Dekker, New York, NY, 2002. E. N. Johnson and A. J. Calise. Neural network adaptive control of systems with input saturation. In Proc. American Control Conf., pages 3527 – 3532, Arlington, VA, June 2001. doi: 10.1109/ACC.2001.946179. E. N. Johnson and A. J. Calise. Limited authority adaptive flight control for reusable launch vehicles. J. Guid. Control Dyn., 26 (6):906 – 913, Nov. – Dec. 2003. doi: 10.2514/2.6934. Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 45 / 35 Backup Slides References References II H. K. Khalil. Nonlinear Systems. Prentice Hall, Upper Saddle River, NJ, 3 edition, 2002. M. V. Kothare, P. J. Campo, M. Morari, and C. N. Nett. A unified framework for the study of anti-windup designs. Automatica, 30(12):1869 – 1883, Dec. 1994. doi: 10.1016/0005-1098(94)90048-5. F. Morabito, A. R. Teel, and L. Zaccarian. Nonlinear antiwindup applied to Euler-Lagrange systems. IEEE Trans. Robot. Autom., 20(3):526 – 537, June 2004. doi: 10.1109/TRA.2004.824933. E. F. Mulder, P. Y. Tiwari, and M. V. Kothare. Simultaneous linear and anti-windup controller synthesis using multiobjective convex optimization. Automatica, 45(3):805 – 811, Mar. 2009. doi: 10.1016/j.automatica.2008.10.019. J. B. Rosen. The gradient projection method for nonlinear programming. part I. linear constraints. J. Soc. Ind. Appl. Math., 8 (1):181 – 217, Mar. 1960. J. B. Rosen. The gradient projection method for nonlinear programming. part II. nonlinear constraints. J. Soc. Ind. Appl. Math., 9(4):514 – 532, Dec. 1961. J. Sofrony, M. C. Turner, I. Postlethwaite, O. Brieger, and D. Leiβling. Anti-windup synthesis for PIO avoidance in an experimental aircraft. In Proc. 45th IEEE Conf. Decision and Control, pages 5412 – 5417, San Diego, CA, Dec. 2006. doi: 10.1109/CDC.2006.377217. M. Soroush and P. Daoutidis. Optimal windup and directionality compensation in input-constrained nonlinear systems. In V. Kapila and K. M. Grigoriadis, editors, Actuator Saturation Control, Control Eng., chapter 9, pages 227 – 246. Marcel Dekker, New York, NY, 2002. G. Stein. Respect the unstable. IEEE Control Syst. Mag., 23(4):12 – 25, Aug. 2003. doi: 10.1109/MCS.2003.1213600. S. Tarbouriech and M. Turner. Anti-windup design: an overview of some recent advances and open problems. IET Control Theory Appl., 3(1):1 – 19, Jan. 2009. doi: 10.1049/iet-cta:20070435. A. Visioli. Practical PID Control. Adv. Ind. Control. Springer, London, United Kingdom, 2006. S.-S. Yoon, J.-K. Park, and T.-W. Yoon. Dynamic anti-windup scheme for feedback linearizable nonlinear control systems with saturating inputs. Automatica, 44(12):3176 – 3180, Dec. 2008. doi: 10.1016/j.automatica.2008.10.003. D. Zhang and A. Nagurney. On the stability of projected dynamical systems. J. Optim. Theory Appl., 85(1):97 – 124, Apr. 1995. doi: 10.1007/BF02192301. Justin Teo (ACL, MIT) Gradient Projection Anti-windup Scheme Dec. 20, 2010 46 / 35