Robust Control: Past Successes and Future...
Transcript of Robust Control: Past Successes and Future...
![Page 1: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/1.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Robust Control: Past Successes and Future Directions
Peter SeilerUniversity of Minnesota
January 22, 2019
![Page 2: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/2.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Outline
• Brief Overview of Robust Control
• Robustness of Time-Varying Systems
• Future Directions
• Conclusions
2
![Page 3: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/3.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Pillars of Robust Control
1. Multivariable Optimal Control
• H2, H∞, DK-synthesis
2. Fundamental Limitations of Control
• Bode sensitivity integral, complementary sensitivity integrals, constraints due to right-half plane poles and zeros.
3. Uncertainty Modeling and Robustness Analysis
• Linear Fractional Transformations (LFTs), Structured Singular Value (m), Integral Quadratic Constraints (IQCs)
3
![Page 4: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/4.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Pillars of Robust Control
1. Multivariable Optimal Control
• H2, H∞, DK-synthesis
2. Fundamental Limitations of Control
• Bode sensitivity integral, complementary sensitivity integrals, constraints due to right-half plane poles and zeros.
3. Uncertainty Modeling and Robustness Analysis
• Linear Fractional Transformations (LFTs), Structured Singular Value (m), Integral Quadratic Constraints (IQCs)
4
![Page 5: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/5.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Many design objectives: Stability, disturbance rejection, reference tracking, noise rejection, moderate actuator commands, adequate robustness margins.
Basic Limitation: S+T=1
Typically require |S| ≪ 1 at low frequencies for reference tracking and disturbance rejection.
Basic Feedback Loop
5
n
d
u
yPC
-r
e
ymeas
-T
S Sensitivity
Complementary
Sensitivity
![Page 6: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/6.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Plant Uncertainty
6
A simplified model P is used for control design.
• Actual dynamics are complex and have part-to-part variation.
• We lose model fidelity as we go to higher frequencies.
Experimental frequency
responses (blue) and
simplified model (black).
Voice Coil Motor
![Page 7: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/7.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Stability Margins: Safety Factors for Control
Consider simple plant uncertainty types:
• Gain: aP where a varies from its nominal anom=1
• Phase: ejq where q varies from its nominal qnom=0
Classical Margins: Largest gain/phase variations that can be tolerated before closed-loop instability occurs.
• Nyquist stability criterion is ideal for studying these margins
• Easily computed in SISO case, e.g. from frequency responses.
Disk Margin: If |S(jw)| is large then a small (combined) gain/phase variation can cause instability.
7
![Page 8: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/8.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Typical Sensitivity Objectives
• Performance: “Small” |S| up to 0 dB bandwidth WS
• Robustness: |S|≤ 2 (=6dB) at all frequencies (No Peaks)
8
Typical sensitivity
response (red)
and design
objectives (black)
![Page 9: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/9.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Suppose P is stable so that C = 0 is a stabilizing controller.
Conservation of Sensitivity
9
AND
![Page 10: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/10.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Suppose P is stable so that C = 0 is a stabilizing controller.
Conservation of Sensitivity
10
AND
Improving sensitivity at some frequencies leads to degradations at others.
This follows from the Bode Sensitivity Integral.
Blue: Sensitivity with C=0
Red: Sensitivity for another
stabilizing controller
![Page 11: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/11.jpg)
AEROSPACE ENGINEERING AND MECHANICS
[1] Bode, Network Analysis and Feedback Amplifier Design, 1945.
[2] Freudenberg and Looze, Frequency Domain Properties of Scalar and Multivariable Feedback
Systems, Springer-Verlag, 1988.
Assume PC has relative degree 2
and S(s) is stable. Then:
where pk are the unstable (RHP) poles of PC.
(Note: )
Bode Integral Theorem [1,2]
11
![Page 12: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/12.jpg)
AEROSPACE ENGINEERING AND MECHANICS
[1] Bode, Network Analysis and Feedback Amplifier Design, 1945.
[2] Freudenberg and Looze, Frequency Domain Properties of Scalar and Multivariable Feedback
Systems, Springer-Verlag, 1988.
[3] Stein, Respect the Unstable, Bode Lecture, 1989 (and IEEE CSM, 2003)
Assume PC has relative degree 2
and S(s) is stable. Then:
where pk are the unstable (RHP) poles of PC.
(Note: )
Bode Integral Theorem [1,2]
12
This a key conserved quantity in feedback design.
Improving performance (e.g. increased bandwidth) comes
at the expense of reduced robustness (peak in |S|) [3].
![Page 13: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/13.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Advanced Robustness Analysis
Move beyond classical SISO stability (gain/phase) margins
1. Multi-loop (MIMO) systems with multiple uncertainties
2. More detailed uncertainty descriptions including
• Parametric,
• Non-parametric (dynamic)
• Nonlinearities, e.g. saturation
3. Consider both robust stability and robust performance
Developments go back to the Lur’e problem (40’s) with key contributions in the 80’s and 90’s:
• m: Safonov, Stein, Doyle, Packard, …
• IQCs: Yakubovich, Megretski, Rantzer, …
13
Integral
Quadratic
Constraints
(IQCs)
Structured
Singular
Value (m)
![Page 14: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/14.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Non-parametric (Dynamic) Uncertainty
Model nominal behavior with LTI system 𝐺0.
Uncertainty modeled by LTI systems ෨𝐺 close to 𝐺0 in frequency response, e.g. small additive error.
14
Error
Bound
![Page 15: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/15.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Uncertainty Modeling
• Unstable plant with uncertain pole and input gain:
• First-order actuator with additive dynamic uncertainty
• Proportional-Integral control
15
Consider SISO feedback system:
![Page 16: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/16.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Uncertainty Modeling
Separate known from the uncertain
16
Uncertainty is
typically very
structured
![Page 17: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/17.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Uncertainty Modeling
Re-center and re-scale to normalize the uncertainties
17
where:
1. D=0 gives nominal behavior
2. Range of modeled uncertainty is
Uncertainty set is structured:
![Page 18: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/18.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Robustness Metrics
Stability Margin:
Worst-case Gain:
Comments:
• System is robustly stable if and only if km >1.
• Both metrics can be converted to a (freq. domain) m test.
• Algorithms compute bounds that provide guarantees on performance and bad instances of uncertainties.
• IQCs extend the framework to include nonlinearities.
18
![Page 19: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/19.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Numerical Algorithms and Software
19
% Unstable plant with parametric uncertainty
a = ureal('a',1, 'Range', [0.8 1.1]);
b = ureal('b',2, 'Range', [1.7 2.6]);
P = tf(b, [1 -a]);
% Actuator with non-parametric (dynamic) unc.
nomAct = tf(10, [1 10]);
DeltaE = ultidyn(‘DeltaE',[1 1]);
A = nomAct + 0.1*DeltaE;
% Uncertain closed-loop (d->e) with PI control
C = tf([3 4.5],[1 0]);
R = feedback(-P, A*C);
% Robust stability and worst-case gain
[StabMargin, DestabilizingUncert] = robstab(R);
[wcGain, OffendingUncertainty] = wcgain(R);
Reliable software to create uncertainty models & perform analyses.
• Matlab’s Robust Control Toolbox (Safonov & Chiang), (Balas, Doyle, Glover, Packard, & Smith), (Gahinet, Nemirovski, Laub, & Chilali)
• ONERA’s Systems Modeling, Analysis and Control Toolbox (Biannic, Burlion, Demourant, Ferreres, Hardier, Loquen, & Roos)
Example Matlab code to
assess robustness of
simple feedback loop.
![Page 20: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/20.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Numerical Algorithms and Software
20
Reliable software to create uncertainty models & perform analyses.
• Matlab’s Robust Control Toolbox (Safonov & Chiang), (Balas, Doyle, Glover, Packard, & Smith), (Gahinet, Nemirovski, Laub, & Chilali)
• ONERA’s Systems Modeling, Analysis and Control Toolbox (Biannic, Burlion, Demourant, Ferreres, Hardier, Loquen, & Roos)
Numerical algorithms continue to be developed, e.g. in Matlab:
• Structured 𝐻∞(R2010b) and systune (R2014a): Based on work by (Gahinet, Apkarian, Noll)
• m without frequency gridding (R2016b): Based on work by (Gahinet, Balas, Packard, Seiler) and (Biannic, Ferreres, Roos)
• Automatic regularization for H2 (R2017b) and 𝐻∞ synthesis (R2018b): Based on work by (Gahinet, Packard, Seiler)
![Page 21: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/21.jpg)
AEROSPACE ENGINEERING AND MECHANICS
(My Recent) Applications of Robust Control
• Wind farm modeling and control (Annoni ‘16, Singh, Hoyt)
• Individual turbine control (Wang ‘16, Ossmann, Theis)
• UAV control with a single aerodynamic surface (Venkataraman ’18)
• Flexible aircraft (Kotikalpudi ’17, Theis ’18, Gupta, Pfifer)
• Dual stage hard disk drives with Seagate (Honda ‘16)(Years refer to Ph.D. theses.)
21
![Page 22: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/22.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Outline
• Brief Overview of Robust Control
• Robustness of Time-Varying Systems
• Joint work with M. Arcak, A. Packard, M. Moore, and C. Meissen at UC, Berkeley.
• Funded by ONR BRC with B. Holm-Hansen at Tech. Monitor
• Future Directions
• Conclusions
22
![Page 23: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/23.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Time-Varying Systems
23
Wind Turbine
Periodic /
Parameter-Varying
Flexible Aircraft
Parameter-Varying
Vega Launcher
Time-Varying
(Source: ESA)
Robotics
Time-Varying
(Source: ReWalk)
Issue: Few numerically reliable methods to assess
the robustness of time-varying systems.
![Page 24: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/24.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Finite Horizon Analysis
Goal: Assess the robustness of linear time-varying (LTV) systems on finite horizons.
Issue: Classical Gain/Phase Margins focus on (infinite horizon) stability and frequency domain concepts.
24
Instead focus on:
• Finite horizon metrics, e.g. induced gains and reachable sets.
• Effect of disturbances and (parametric, non-parametric and nonlinear) uncertainties.
• Time-domain analysis conditions.
![Page 25: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/25.jpg)
AEROSPACE ENGINEERING AND MECHANICS
(Nominal) Finite Horizon Analysis
25
Nominal LTV System
Analysis Objective
![Page 26: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/26.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Nominal Analysis with Dissipation Inequalities
26
Theorem [1,2]
[1] Moore, Finite Horizon Robustness Analysis using IQCs, MS Thesis, Berkeley, 2015.
[2] Willems, Dissipative Dynamical Systems: Parts i and ii, 1972.
Proof
![Page 27: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/27.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Nominal Analysis with Dissipation Inequalities
27
Theorem [1,2]
[1] Moore, Finite Horizon Robustness Analysis using IQCs, MS Thesis, Berkeley, 2015.
[2] Willems, Dissipative Dynamical Systems: Parts i and ii, 1972.
![Page 28: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/28.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Nominal Analysis with Dissipation Inequalities
28
Theorem [1,2]
[1] Moore, Finite Horizon Robustness Analysis using IQCs, MS Thesis, Berkeley, 2015.
[2] Willems, Dissipative Dynamical Systems: Parts i and ii, 1972.
[3] Green & Limebeer, Linear Robust Control, 1995.
[4] Iannelli, Seiler, Marcos, “Construction of worst-case disturbances for LTV systems…”, 2019.
Comments• The dissipation inequality is equivalent to Riccati conditions [3] but enables extensions to robustness analysis.
• Numerically reliable algorithm to construct worst-case disturbance [4].
![Page 29: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/29.jpg)
AEROSPACE ENGINEERING AND MECHANICS
(Robust) Finite-Horizon Analysis
29
Uncertain LTV System
Analysis Objective
![Page 30: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/30.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Integral Quadratic Constraints (IQCs) [1,2]
30
[1] Yakubovich, S-procedure in nonlinear control theory, 1971.
[2] Megretski and Rantzer, System Analysis via Integral Quadratic Constraints, TAC, 1997.
Definition
Notationally we consider only a single block. If D has block
structure then we can specify different IQCs for each block.
![Page 31: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/31.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Example: Sector-bounded Nonlinearity
31
![Page 32: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/32.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Example: Sector-bounded Nonlinearity
32
![Page 33: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/33.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Example: Non-parametric (Dynamic) Uncertainty
33
![Page 34: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/34.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Example: Non-parametric (Dynamic) Uncertainty
34
Frequency-
Domain IQC
Spectral
Factorization
![Page 35: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/35.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Example: Non-parametric (Dynamic) Uncertainty
35
Frequency-
Domain IQC
[1] Balakrishnan, Lyapunov Functionals in Complex m Analysis, TAC, 2002.
(Invoke Causality)
![Page 36: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/36.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Additional IQC Details
• Most IQCs are related to previous robust stability results
• IQC for sector nonlinearities related to the circle criterion
• IQC for LTI uncertainties related to D-scales in m analysis
• A dictionary of additional IQC for various uncertainties / nonlinearities is given in [1].
• IQCs for passive operators, static memoryless nonlinearities (Popov, Zames-Falb), time-delays, real parameters, etc.
• Many IQCs are specified in the frequency domain
• A technical J-spectral factorization result can be used to convert freq. domain IQCs into time-domain IQCs [2,3].
36
[1] Megretski & Rantzer, System analysis via IQCs, TAC, 1997. [IQCs derived based on much prior literature]
[2] Seiler, Stability Analysis with Dissipation Inequalities and Integral Quadratic Constraints, TAC, 2015.
[3] Hu, Lacerda, Seiler, Robustness Analysis of Uncertain Discrete-Time System with … IQCs, IJRNC, 2016.
![Page 37: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/37.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Robustness Analysis
37
![Page 38: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/38.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Robust Finite Horizon Analysis
38
Theorem [1,2]
[1] Moore, Finite Horizon Robustness Analysis using IQCs, MS Thesis, Berkeley, 2015.
[2] Seiler, Moore, Meissen, Arcak, Packard, Finite Horizon Robustness Analysis of LTV Systems Using
IQCs, arXiv 2018 and Automatica 2019.
Proof
![Page 39: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/39.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Robust Finite Horizon Analysis
39
Theorem [1,2]
[1] Moore, Finite Horizon Robustness Analysis using IQCs, MS Thesis, Berkeley, 2015.
[2] Seiler, Moore, Meissen, Arcak, Packard, Finite Horizon Robustness Analysis of LTV Systems Using
IQCs, arXiv 2018 and Automatica 2019.
![Page 40: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/40.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Numerical Algorithms and Software
• Robustness Algorithms
• Differential LMI can be “solved” via convex optimization using basis functions for 𝑃 ∙ and gridding on time [1].
• A more efficient algorithm mixes the differential LMI and a related Riccati Differential Equation condition [2].
• Similar methods developed for LPV [4,5] and periodic systems [6].
• LTVTools Software [3]
• Time-varying state space system objects, e.g. obtained from Simulink snapshot linearizations.
• Includes functions for nominal and robustness analyses.
40
[1] Moore, Finite Horizon Robustness Analysis using IQCs, MS Thesis, Berkeley, 2015.
[2] Seiler, Moore, Meissen, Arcak, Packard, Finite Horizon Robustness Analysis of LTV Systems Using IQCs,
arXiv 2018 and Automatica 2019.
[3] https://z.umn.edu/LTVTools
[4] Pfifer & Seiler, Less Conservative Robustness Analysis of LPV Systems Using IQCs, IJRNC, 2016.
[5] Hjartarson, Packard, Seiler, LPVTools: A Toolbox for Modeling, Analysis, & Synthesis of LPV Systems, 2015.
[6] Fry, Farhood, Seiler, IQC-based robustness analysis of discrete-time LTV systems, IJRNC 2017.
![Page 41: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/41.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Two-Link Robot Arm
41
Nonlinear dynamics [MZS]:ሶ𝜂 = 𝑓(𝜂, 𝜏, 𝑑)
where
𝜂 = 𝜃1, ሶ𝜃1, 𝜃2, ሶ𝜃2𝑇
𝜏 = 𝜏1, 𝜏2𝑇
𝑑 = 𝑑1, 𝑑2𝑇
t and d are control torques and disturbances at the link joints.
[1] R. Murray, Z. Li, and S. Sastry. A Mathematical Introduction to Robot Manipulation, 1994.
Two-Link Diagram [1]
![Page 42: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/42.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Nominal Trajectory in Cartesian Coordinates
42
![Page 43: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/43.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Analysis
43
Nonlinear dynamics:ሶ𝜂 = 𝑓(𝜂, 𝜏, 𝑑)
Linearize along the finite –horizon trajectory ҧ𝜂, ҧ𝜏, 𝑑 = 0ሶ𝑥 = 𝐴 𝑡 𝑥 + 𝐵 𝑡 𝑢 + 𝐵 𝑡 𝑑
Design finite-horizon state-feedback LQR gain.
Goal: Compute bound on the final position accounting for disturbances and LTI uncertainty D at 2nd joint.
![Page 44: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/44.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Monte-Carlo Simulations
44
LTV simulations with randomly sampled disturbances and uncertainties (overlaid on nominal trajectory).
![Page 45: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/45.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Robustness Bound
45
Cyan disk is bound computed in 102 sec using IQC/DI method
Bound accounts for disturbances ‖d‖≤5 and ‖Δ‖≤0.8
![Page 46: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/46.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Worst-Case Uncertainty / Disturbance
46
Randomly sample D to find “bad” perturbation and compute corresponding worst-case disturbance using method in [1]
[1] Iannelli, Seiler, Marcos, Construction of worst-case disturbances for LTV systems…, 2019.
![Page 47: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/47.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Closed-Loop Robust L2-to-Euclidean Gain
Two Controllers:
• Finite-Horizon LQR with state feedback
• Output Feedback using high pass filter 𝝉s/(𝝉s+1) to estimate angular rates
47
Finite horizon robustness is degraded by output feedback with rate estimates.
![Page 48: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/48.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Closed-Loop Robust L2-to-Euclidean Gain
Two Controllers:
• Finite-Horizon LQR with state feedback
• Output Feedback using high pass filter 𝝉s/(𝝉s+1) to estimate angular rates
48
Finite horizon robustness is degraded by output feedback with rate estimates.
0.1
![Page 49: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/49.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Closed-Loop Robust L2-to-Euclidean Gain
Two Controllers:
• Finite-Horizon LQR with state feedback
• Output Feedback using high pass filter 𝝉s/(𝝉s+1) to estimate angular rates
49
Finite horizon robustness is degraded by output feedback with rate estimates.
![Page 50: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/50.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Impact of Using High Pass Rate Estimator
50
![Page 51: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/51.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Outline
• Brief Overview of Robust Control
• Robustness of Time-Varying Systems
• Future Directions
• Robustness in Reinforcement Learning
• Design for Optimization Algorithms
• Conclusions
51
![Page 52: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/52.jpg)
AEROSPACE ENGINEERING AND MECHANICS
“Model-Free” Reinforcement Learning
• Goal: Train a control policy from data to maximize a cumulative reward
• Training data obtained from a simulator or the real system
• Often assume state feedback
• Many algorithms (Q-learning, value iteration, policy iteration, policy search) [1,2,3]
• Algorithms have close connections to dynamic programming and optimal control.
52
[1] D.P. Bertsekas, “Reinforcement Learning and Optimal Control,” 2019.
[2] R.S. Sutton and A.G. Barto, “Reinforcement Learning: An Introduction,” 2018.
[3] C. Szepesvári, “Algorithms for Reinforcement Learning,” 2010.
![Page 53: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/53.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Is Robustness an Issue in RL?
Training via simulation
53
Training on real system
![Page 54: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/54.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Is Robustness an Issue in RL?
Training via simulation
• Training can exploit flaws in the simulator [1].
• Loss of performance transitioning from simulator to real system.
54
Robotic Walking in MuJoCo
[1] B. Recht, “A Tour of Reinforcement Learning,” arXiv, 2018.
Training on real system
![Page 55: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/55.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Is Robustness an Issue in RL?
Training via simulation
• Training can exploit flaws in the simulator [1].
• Loss of performance transitioning from simulator to real system.
55
Robotic Walking in MuJoCo
[1] B. Recht, “A Tour of Reinforcement Learning,” arXiv, 2018.
Training on real system
• Part to part variation (train on one system and implement on many)
• Changes in system dynamics over time (temperature dependence, environmental effects, etc….)
![Page 56: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/56.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Summary: Recovering Robustness in RL
• Robustness issues can arise in output-feedback controllers trained by RL [2]
• Linear Quadratic Gaussian (LQG) Control can be solved via RL
• A well-known counterexample by Doyle [1] demonstrates that LQG controllers can have arbitrarily small margins.
56
[1] J. Doyle. Guaranteed margins for LQG regulators, IEEE TAC, 1978.
[2] Venkataraman & Seiler, Recovering Robustness in Model-Free Reinforcement Learning, ’18 arXiv and ‘19 ACC submission.
![Page 57: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/57.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Summary: Recovering Robustness in RL
• Robustness issues can arise in output-feedback controllers trained by RL [2]
• Linear Quadratic Gaussian (LQG) Control can be solved via RL
• A well-known counterexample by Doyle [1] demonstrates that LQG controllers can have arbitrarily small margins.
57
[1] J. Doyle. Guaranteed margins for LQG regulators, IEEE TAC, 1978.
[2] Venkataraman & Seiler, Recovering Robustness in Model-Free Reinforcement Learning, ’18 arXiv and ‘19 ACC submission.
• Robustness can be recovered by introducing (synthetic) input perturbations during the RL training [2].
![Page 58: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/58.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Linear Quadratic Gaussian (LQG)
Minimize
Subject To:
The optimal controller has an observer/state-feedback form
Gains (K,L) computed by solving two Riccati equations.
This solution is model-based, i.e. it uses data A,B,C, etc
58
![Page 59: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/59.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Reinforcement Learning
• Partially Observable Markov Decision Processes (POMDPs)
• Set of states, S
• Set of actions, A
• Reward function, r: S x A → ℝ
• State transition probability, T
• Set of observations and observation probability, O
• Many methods to synthesize a control policy from input/output data to maximize the cumulative reward
• The LQG problem is a special case of this RL formulation
59
![Page 60: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/60.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Doyle’s Example (‘78 TAC)
• LQR state-feedback regulators have provably good margins.
• Doyle’s example shows that LQG regulators can have arbitrarily small input margins.
60
![Page 61: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/61.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Doyle’s Example (‘78 TAC)
• LQR state-feedback regulators have provably good margins.
• Doyle’s example shows that LQG regulators can have arbitrarily small input margins.
• Doyle’s example can also be solved within RL framework using direct policy search:
• RL will converge to the optimal LQG control with infinite data collection. Thus RL can also have poor margins.
61
![Page 62: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/62.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Recovering Robustness
• Increase process noise during training?
• This causes margins to decrease on Doyle’s example
• Process noise is not model uncertainty
62
![Page 63: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/63.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Recovering Robustness
• Increase process noise during training?
• Modify reward to increase state penalty or decrease control penalty?
• Again, this causes margins to decrease on Doyle’s example
• Trading performance vs. robustness via the reward function can be difficult or counter-intuitive
63
![Page 64: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/64.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Recovering Robustness
• Increase process noise during training?
• Modify reward to increase state penalty or decrease control penalty?
• Inject synthetic gain/phase variations at the plant input (and output?) during the training phase.
D=1+d where
d is U[-b,b]
64
![Page 65: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/65.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Results On Doyle’s Example
65
![Page 66: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/66.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Results on Simplified Flexible System
• Model has 4-states (Rigid body and lightly damped modes)
• RL applied to 3-state controller parameterization• LQG controller is not in the control policy parameterization
• Still converges to policy with small margins
• Robustness recovered with synthetic perturbations during training
66
![Page 67: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/67.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Next Steps
• How should synthetic perturbations be introduced during training?
• Can we make any rigorous claims about the proposed method? Performance certification?
• Impact of fundamental performance limits on RL policies?
67
![Page 68: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/68.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Outline
• Brief Overview of Robust Control
• Robustness of Time-Varying Systems
• Future Directions
• Robustness in Reinforcement Learning
• Design for Optimization Algorithms
• Conclusions
68
![Page 69: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/69.jpg)
AEROSPACE ENGINEERING AND MECHANICS
First-order optimization algorithms as robust control
Assumptions on 𝑓 (uncertain plant)• Strongly convex (𝑚)• Lipschitz gradients (𝐿)
Algorithm (controller)• Finite-dim, strictly proper, linear system• input: gradient at iterate• output: next iterate
Automated Analysis with IQC/SDP• characterize ∇𝑓 with IQCs • certify convergence-rate of interconnection
Extensions• Gradient noise• Constrained optimization• Algorithm design• Stochastic Gradient Descent
69
Lessard, Hu, Recht,
Seiler, Rantzer, Packard
![Page 70: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/70.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Outline
• Brief Overview of Robust Control
• Robustness of Time-Varying Systems
• Future Directions
• Conclusions
70
![Page 71: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/71.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Conclusions
• Robust control has a long history with many successes
1. Multivariable Optimal Control
2. Fundamental Limitations of Control
3. Uncertainty Modeling and Robustness Analysis
• Robust control techniques can solve emerging problems
1. Robustness in controls designed via data-driven (RL) methods
2. Design and analysis of optimization algorithms
• Acknowledgements:• Funding: NSF, AFOSR, ONR, NASA, Seagate, MSI, Xcel RDF, MnDrive
• Past PhDs & Visitors: Annoni, Hu, Honda, Kotikalpudi, Lacerda, Ossmann, Peni, Pfifer, Takarics, Theis, Venkataraman, Wang
https://www.aem.umn.edu/~SeilerControl/
71
![Page 72: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/72.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Backup Slides
![Page 73: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/73.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Typical S+T=1 Tradeoff
73
Large loop gain |PC|:
Good reference tracking
Poor noise rejection
Small loop gain |PC|:
Poor reference tracking
Good noise rejection
![Page 74: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/74.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Typical S+T=1 Tradeoff
74
Large loop gain |PC|:
Good reference tracking
Poor noise rejection
Small loop gain |PC|:
Poor reference tracking
Good noise rejection
Crossover Region:
Poor reference tracking
AND Poor noise rejection
![Page 75: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/75.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Bode Integral Theorem and “Peaking”
A procedure to avoid peaking could be:
• Obtain significant Sensitivity reduction over [0, WS].
This incurs a large negative integral which must be balanced.
• Maintain |S(jω)| slightly larger than 1 over a wide interval.
This incurs a positive integral balancing the negative integral.
• Make |PC| approach 0 quickly at higher frequencies so that |S| quickly approaches 1.
75
Accumulate a large area by having |S| just
exceed 1 over a large frequency range.
0
![Page 76: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/76.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Available Bandwidth
The Bode Integral theorem may appear to be a minor constraint, e.g. spreading area over a large frequency band.
Stein (‘89 Bode Lecture, ’03 CSM):
a key fact about physical systems is that they do not exhibit good frequency response fidelity beyond a certain bandwidth. … Let us call that bandwidth the “available bandwidth,” Wa
76
Accumulate a large area by having |S| just
exceed 1 over a large frequency range.
0
![Page 77: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/77.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Available Bandwidth
The Bode Integral theorem may appear to be a minor constraint, e.g. spreading area over a large frequency band.
Stein (‘89 Bode Lecture, ’03 CSM):
a key fact about physical systems is that they do not exhibit good frequency response fidelity beyond a certain bandwidth. … Let us call that bandwidth the “available bandwidth,” Wa
77
The available bandwidth due to physical (hardware) constraints requires positive area be accumulated over a finite frequency band.
Consequence: Improving performance (e.g. increased bandwidth) comes at the expense of reduced robustness (peak in |S|).
![Page 78: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/78.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Consequence of Available Bandwidth
|PC| must roll-off quickly above Wa
78
0
roughly
Performance is constrained by the Bode integral
and robustness requirements.
![Page 79: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/79.jpg)
AEROSPACE ENGINEERING AND MECHANICS
Partial Dictionary of IQCs [1]
Uncertainty
1. Passive
2. Norm-bounded LTI
3. Constant Real Parameter
4. Varying Real Parameter
5. Unit Saturation
79
[1] Megretski & Rantzer, System analysis via IQCs, TAC, 1997. [IQCs derived based on much prior literature]
IQC Multiplier
![Page 80: Robust Control: Past Successes and Future DirectionsSeilerControl/Papers/Slides/2019/Jan22_RobustControl... · Pillars of Robust Control 1. Multivariable Optimal Control • H 2,](https://reader034.fdocuments.net/reader034/viewer/2022042103/5e815481d4fb876fb84b41d6/html5/thumbnails/80.jpg)
AEROSPACE ENGINEERING AND MECHANICS
LTV Toolchain
80
Simulink Model of
Robotic Arm
% Matlab snapshot linearizations
% along nominal trajectoryio(1) = linio('TwoLinkRobotOL/Input Torque',1,'input');
io(2) = linio('TwoLinkRobotOL/Two Link Robot Arm',1,'output');
sys = linearize('TwoLinkRobotOL',io,Tgrid);
% Construction of LTV ModelG = tvss(sys,Tgrid);