BTEC 2018

Engineering Our Future

Innovation to change the world

St. Charles,
Philadelphia
Missouri
July
May8—12,
7-10, 2019
2018

Robust Adaptive Control

w/ applications in flight control ★

Dr. Heather Hussain

Guidance, Navigation, Controls, and Autonomy (GNC&A)
The Boeing Company, Tukwila, WA
[email protected]

sponsored by the Boeing Strategic University Initiative 1

 American Control Conference 2019

Abstract
Heather Hussain
With the advent of each next generation technology, demands for a rapidly reconfigurable control
system yielding invariant performance under increasingly unknown or widely varying operating
conditions becomes crucial. Adaptive control has long been viewed as one such control method, with
implementation on high performance aerial vehicles providing nearly uniform performance across the
flight envelope even with limited a priori knowledge on the aircraft’s aerodynamic characteristics. This
adaptation to parametric uncertainties is achieved through a process of online measurement,
evaluation, and compensation through the control input.
While the foundations of robust adaptive control theory were laid in the early 1980’s, obtaining
quantifiable and practically meaningful robust stability margins for adaptive systems remained an open
problem. Successful implementation of adaptive control theory as a viable control solution can only be
achieved when global robustness properties, especially with respect to unmodeled dynamics, are well
understood. This research proposes a solution to this long standing open problem for a class of linear
time-invariant plants, whose states are accessible.
With the use of a modified adaptive update law and sufficient conditions of a frequency-domain
criterion, it is shown that the underlying closed-loop system has globally bounded solutions. That is,
the overall adaptive system is shown to have analytically computable robustness margins that hold for
arbitrary initial conditions.
It is also shown that, with these global properties established, specific conditions can be derived under
which the advantage of adaptation over non-adaptive solutions for the control of uncertain systems is
made clear. This advantage lies in the fact that parameter adaptation allows learning of the
uncertainties whenever the effect of unmodeled dynamics is small, leading to small tracking errors and
improved robustness margins.

2
 American Control Conference 2019

Presenter

HUSSAIN, HEATHER Heather Hussain received the B.S. degree and M.S. degree in mechanical
engineering from the Rochester Institute of Technology, Rochester, NY,
USA, in 2012, and the Sc.D. degree in mechanical engineering at the
Massachusetts Institute of Technology (MIT), Cambridge, MA, USA in
2017.
Her work experience comprises several internships spanning the
aerospace and consumer electronics industries– namely, in Product
Design at Apple Inc., as a research Scholar at the Munitions Directorate
of the Air Force Research Laboratory, and her work in the design and
development of verifiable adaptive flight control systems at The Boeing
Company.
Ms. Hussain’s doctoral research at MIT was sponsored by the Boeing
Strategic University Initiative under the direction of Dr. Eugene
Lavretsky and Dr. Anuradha Annaswamy. Ms. Hussain joined BR&T’s
Guidance, Navigation, Control, and Autonomy (GNC&A) group in
September 2017. Her research interests lie in adaptive control theory,
particularly with applications in aerospace.
Ms. Hussain is a member of AIAA and IEEE.
 American Control Conference 2019

Outline
▫ Introduction/Motivating Example
▫ Need for Robust Adaptive Control
▫ Significant Earlier Work
▫ Main Result: Computable Robustness Margins for Adaptive Systems
▫ Main Idea & Key Elements
▫ “Tutorial” using Numerical Example
▫ Validation via Simulation Studies
▫ Why Adapt?
▫ Operator Equivalence
▫ Persistence of Excitation & Parameter Convergence
▫ Summary
▫ Limitations & Future Work

4
 American Control Conference 2019

Direct Model Reference Adaptive Control (MRAC)

Introduction

▫ Adaptive control theory is a mature control discipline that allows for real-time compensation
of uncertainties and changes in system dynamics

▫ Premise: Adapt system parameters to provide a vehicle response that more closely follows
the reference model
▫ Graceful degradation in presence of uncertainties
▫ Ability to continue mission

▫ Challenge: Gains are bounded nonlinear integral paths à closed loop dynamics are
inherently nonlinear
– Gain and Phase margin are not defined during adaptation

5
 American Control Conference 2019
Zq Z e Z↵
V0 ⇠ V0 ⇠ V0 ⇠ 0, M↵ = !p2 , Mq = 2⇣p !p , M e
= 1, (⇣p , !p , kp ) = ( 1, 1, 1), (⇣⌘ , !⌘ , k⌘ ) = (0.89, 17, 172 ), (⇣m , !m , km ) = (1, 3, 1)

Motivation
Adaptive Control
Elevator
Actuator Dynamics
V !n2 e
Reference Model e

s2 + 2⇣!n s + !n2
2 Gm (s) ↵ref

1Elevator Plant
Actuator Dynamics
" # " #" # "Z #
r(t) + P ecmd 0 !n2 e ↵˙ Z↵
V0
1 + ZV0q ↵ V0
e
↵ + P e
= + ⇤ e
s2 + 2⇣!n s + !n2 q̇ M↵ Mq q M
+ -1 e

Feedback Gain
-2 50% (⇤ = 0.5)
0 20✓ 40 60 80 100 120 140 160
Mq = M q + Mq
2

–ref (deg) 1 Non-Adaptive Controller M↵ = M ↵ + M↵

10
– (deg) 05
- 10
-5
-2
-100 not robust
20 to40parametric
60 uncertainties
80 100 120 140 160
�����������
0 20 40 60 80 100 120 140 160
2
1
0
-1
robust to non-parametric uncertainties
-2
0 20 40 60 80 100 120 140 160

Propose an adaptive controller that improves robustness to parametric

uncertainties
✓˙i = i e i xp

Online Estimate
Adaptive Aircraft
Controller Dynamics
Control Input
e

6
 American Control Conference 2019
Zq Z e Z↵
V0 ⇠ V0 ⇠ V0 ⇠ 0, M↵ = !p2 , Mq = 2⇣p !p , M e
= 1, (⇣p , !p , kp ) = ( 1, 1, 1), (⇣⌘ , !⌘ , k⌘ ) = (0.89, 17, 172 ), (⇣m , !m , km ) = (1, 3, 1)

Motivation
Robust Adaptive Control
Elevator
Actuator Dynamics
V !n2 e
Reference Model e

s2 + 2⇣!n s + !n2
Gm (s) ↵ref
Elevator
Actuator Dynamics Plant
" # " #" # "Z #
r(t) + P ecmd !n2 e ↵˙ Z↵
V0
1 + ZV0q ↵ V0
e
↵ + P e
= + ⇤ e
s2 + 2⇣!n s + !n2 q̇ M↵ Mq q M e
+
Feedback Gain
50% (⇤ = 0.5)
✓
Mq = M q + Mq

–ref (deg)
2 Classical Adaptive Controller M↵ = M ↵ + M↵

– (deg) 1

-1
robust to parametric uncertainties
-2
0 20 40 60 80 100 120 140 160
2
12
01
- 10
not robust to non-parametric uncertainties
--21
0 20 40 60 80 100 120 140 160
����������� -2
Goal: Find a controller
2
0 20
that adapts
40
to
60
parametric
80
uncertainties
100 120
and
140
is robust
160
to unmodeled
dynamics – both are inevitable in real-world systems
X
1
0
2 Robust & Adaptive Flight Controller
1
-1
0
-2
0 20 40 60 80 100 120 140 160
-1

-2
0 20 40 60 80 100 120 140 160

7
10
5
0
-5
 American Control Conference 2019

Background
Milestones

“An understanding of fundamental limitations is an essential element in all engineering.”

–Gunter Stein, 1989 Bode Lecture

Model Reference
Adaptive Lyapunov Theory
Control [1, 2] Analyze & Improve
Based Design [3, 4] Performance
Wide Class of Properties
Supersonic flight,
Advances in
MRAC Schemes
post–Sputnik era &
digital computers
Feedback Control Developed [5–8] Nonlinear Widely used in
Industry,
sparked inspiration. Theory & Adaptive +renewed interest
Stability Control in aerospace
MIT Rule applications

1950 1960 1970 1980 1990 2000 2010 2020

Global
Non-robust boundedness to
Behavior of Rohrs’ Various bounded
Example [10] robustness disturbances [11]
Adaptive
modifications
Control
developed to
Uncovered [5, 9, 10] counteract Guaranteed
instabilities [9, 11–14] Robustness Margins
for Adaptive Control
Robust Adaptive Control Systems [15–18]
Main Result

[1] H. P. Whitaker, J. Yamron, A. Kezer, MIT, Design of model-reference adaptive control systems for aircraft. MIT, Instrumentation Laboratory : Jackson & Moreland, 1958. [2] P. V. Osburn, H. P. Whitaker, and A. Kezer, “New developments in the design of model reference adaptive control systems,”
IAS 29th Annual Meeting, no. Paper 61-39. New York, NY: Institute of the Aerospace Sciences, Jan 1961. [3] B. Shackcloth and R. L. Butchart, Synthesis of Model Reference Adaptive Systems by Liapunov’s Second Method. Boston, MA: Springer US, 1965, pp. 145–152. [4] P. Parks, “Liapunov redesign of
model reference adaptive control systems,” IEEE TAC, vol. 11, no. 3, pp. 362–367, Jul 1966. [5] B. Egardt, Stability of adaptive controllers. New York: Springer Verlag, 1979. [6] K. S. Narendra, Y. H. Lin, and L. S. Valavani, “Stable adaptive controller design - part II: proof of stability,” IEEE TAC, vol.
25, pp. 440–448, 1980. [7] I. D. Landau, Adaptive control: the model reference approach. Marcel Dekker, 1979. [8] A. Morse, “Global stability of parameter-adaptive control systems,” IEEE TAC, vol. 25, no. 3, pp. 433–439, Jun 1980. [9] P. A. Ioannou and P. V. Kokotovic, Adaptive Systems with Reduced
Models. New York: Springer–Verlag, 1983. [10] C.Rohrs, L.Valavani, M.Athans, and G.Stein,“Robustness of continuous-time adaptive control algorithms in the presence of unmodeled dynamics,” IEEE TAC, vol. 30, no. 9, pp. 881 – 889, Sep. 1985. [11] K. S. Narendra and A. M. Annaswamy, “Robust
adaptive control in the presence of bounded disturbances,” IEEE TAC, vol. 31, pp. 306–315, 1986. [12] P. A. Ioannou and P. V. Kokotovic, Adaptive Systems with Reduced Models. Secaucus, NJ, USA: Springer-Verlag New York, Inc., 1983. [13] B. Peterson and K. Narendra, “Bounded error adaptive
control,” IEEE TAC, vol. 27, no. 6, 1982. [14] S. Naik, P. Kumar, and B. Ydstie, “Robust continuous-time adaptive control by parameter projection,” IEEE TAC, vol. 37, no. 2, pp. 182 –197, feb 1992. [15] M. Matsutani, “Robust adaptive flight control systems in the presence of time delay,” Ph.D.
dissertation, MIT, Feb 2013.

8
 Ch03-H6927.tex 30/6/2007 12: 19 Page 57

Static Equilibrium and Trim 57

American Control Conference 2019
noticeably large fins and in some cases the aircraft have two fins attached to the outer
edges of the upper fuselage.

Robust Adaptive Control

3.6 CALCULATION OF AIRCRAFT TRIM CONDITION

As described in Section 3.1, the condition for an aircraft to remain in steady trimmed
flight requires that the forces and moments acting on the aircraft sum to zero and that

Problem Statement
it is stable. Thus, in order to calculate the trim condition of an aircraft it is convenient
to assume straight or symmetric flight and to apply the principles described earlier
in Chapter 3. For a given aircraft mass, cg position, altitude and airspeed, symmetric
trim is described by the aerodynamic operating condition, namely angle of attack,
thrust, pitch attitude, elevator angle and flight path angle. Other operating condition
parameters can then be derived as required.

▫ Adaptive control needs to be:

The forces and moments acting on an aeroplane in the general case of steady
symmetric climbing flight are shown in Fig. 3.20 where the symbols have their usual
meanings. Since the aircraft is symmetric, the lateral–directional forces and moments
are assumed to remain in equilibrium throughout, and the problem reduces to the

▫ stable with parametric uncertainties

establishment of longitudinal equilibrium only. Thus, the reference axes are aircraft
body axes which define the plane of symmetry oxz, with the origin o located at the
aircraft cg as shown.

▫ robust to non-parametric uncertainties 3.6.1 Defining the trim condition

The total axial force X is given by resolving the total lift L, total drag D, weight mg
and thrust τ e into the ox axis and these components must sum to zero in trim. Whence

X = L sin αe + τe cos κ − D cos αe − mg sin (αe + γe ) = 0 (3.36)

where αe is the equilibrium body incidence, γe is the steady flight path angle and κ is

X
the inclination of the thrust line to the ox body axis (positive nose up). Similarly, the
total normal force Z is given by resolving the forces into the oz axis and these also
must sum to zero in trim. Whence

Z = mg cos (αe + γe ) − L cos αe − D sin αe − τe sin κ = 0 (3.37)

x
L
M ae V0
te qe ge
Horizon o
D k
zt

z
mg

Figure 3.20 Symmetric forces and moments acting on a trimmed aircraft.

Main Idea: Robustness Margins for Adaptive Systems

▫ Significant earlier work[1-4] is conservative.

– Global boundedness for a narrow class of unmodeled dynamics[1-3]
– Semi-global boundedness for a slightly larger class of unmodeled dynamics[3,4]

1. S.M. Naik, P.R. Kumar, and B.E. Ydstie. Robust Continuous-time Adaptive Control by Parameter Projection. IEEE Transactions on Automatic Control, Feb 1992.
2. M. Matsutani, A.M. Annaswamy, T. Gibson, and E. Lavretsky. Trustable Autonomous Systems using Adaptive Control. In Proceedings of IEEE Conference on Decision and Control, 2011.
3. P. A. Ioannou and J. Sun. Robust Adaptive Control. Prentice Hall, 1996.
4. K. S. Narendra and A. M. Annaswamy. Stable Adaptive Systems. Prentice Hall, 1989.

9
 American Control Conference 2019

Why is this important?

▫ Non-parametric uncertainties are inevitable è need robust control solutions

Uncertainty
Parametric Nonparametric (⌘)
— Unknown stiffness, — Flexible Effects,
Control effectiveness Actuator Dynamics

✓˙i = i e i xp

Adaptive Parameter Estimate Aircraft

Controller Control Input Dynamics
u

Online Information
Theoretically verifiable
robustness margins

Classical Adaptation Classical Adaptation[1] My Adaptation[2]

No Unmodeled Dynamics with Unmodeled Dynamics with Unmodeled Dynamics
Performance (error)

Performance (error)

Performance (error)
Unbounded parameter "
z }| {
drift in the presence of |e|max

|e| ! 1 7 unmodeled dynamics[1,2].

|✓| ! 1 7

|e|desired
|e| ! 0 X
0
Learning ✓max ✓= 1.7= 16.7
Learning Learning max

Bounded
Learning
1. C.E. Rohrs, L. Valavani, Athans, M., and G. Stein. Robustness of Continuous-time Adaptive Control Algorithms in the Presence of Unmodeled Dynamics. IEEE TAC Automatic Control, 1985
2. H. Hussain, M. Matsutani, A. Annaswamy, and E. Lavretsky, “A new approach to robust adaptive control”, ACC 2016

10
 American Control Conference 2019

Why has this been an open problem for so long?

Uncertainty
Parametric Nonparametric (⌘)
— Unknown stiffness, — Flexible Effects,
Control effectiveness Actuator Dynamics

✓˙i = i e i xp

Adaptive Parameter Estimate Aircraft

Controller Control Input Dynamics
u

Online Information

Z
With Parametric Uncertainty: ✓(t) = xp (⌧ )b>
m P e(⌧ )d⌧ ⇒ Nonlinear Time-Varying Closed-Loop
Actuator Dynamics
(s2 + 2⇣!n s)
s2 + 2⇣!n + !n2

⌘ = G (s)u
+
⇒ State-Dependent Disturbance
V + P La p
With Nonparametric Uncertainty:
a a
(s Lp )
Plant

Extremely difficult to show global

boundedness
 American Control Conference 2019

Main Idea
Robustness to Unmodeled Dynamics

Adaptive Law + Projection + Provably Correct Learning Bounds =

Main Result: Analytical Guarantees for global
boundedness of projection-based adaptive systems with
unmodeled dynamics.
H. Hussain, M. Matsutani, A. Annaswamy, and E. Lavretsky, Adaptive Control of Scalar Plants in the Presence of Unmodeled Dynamics, IFAC ALCOSP, July 2013.
A. Annaswamy, T. Gibson, H. Hussain, and E. Lavretsky. Practical Adaptive Control, 16th Yale Workshop, June 2013.
H. Hussain, A. Annaswamy, and E. Lavretsky, “Robust Adaptive Control in the Presence of Unmodeled Dynamics: A Counter to Rohrs’s Counterexample,” AIAA Guidance, Navigation, and Control Conference, Aug 2013.
H. Hussain, M. Matsutani, A. Annaswamy, and E. Lavretsky, “Computable Delay Margins for Adaptive Systems with State Variables Accessible,” IEEE TAC (To Appear).
H. Hussain, A. Annaswamy, and E. Lavretsky, “A New Approach to Robust Adaptive Control,” ACC, July 2016.
H. Hussain, A. Annaswamy, and E. Lavretsky, “Adaptive control of second-order plants in the presence of unmodeled dynamics,” IFAC ALCOSP, June 2016.
H. Hussain, C. Sharma-Subedi, A. Annaswamy, and E. Lavretsky, “Robustness of Adaptive Control Systems to Unmodeled Dynamics: A Describing Function Viewpoint,” AIAA GNC, January 2017

12
 American Control Conference 2019

Main Result
Robustness to Unmodeled Dynamics

Theorem 1(a). Consider the adaptive system described by Theorem 1(b). Suppose r(t) = 0 8t.
ẋ = Aol x + B⇤(t)y + bol r ẋ = (A0 + Ã(t))x + B1 ⇤1 (t)y1
y = C> x . y1 = C1> x .
Gol (s) , C> (sI Aol ) B G1 (s) , C1> (sI A0 ) B1
1 1

If the transfer function matrix If the transfer function matrix

Z(s) , (I + ⇤max Gol (s))(I ⇤max Gol (s)) Z1 (s) , (I + ⇤1,max G1 (s))(I ⇤1,max G1 (s))
1 1

is strictly positive real with is strictly positive real with

⇤max = diag(#0,max , #1,max , · · · , #n 1,max ) ⇤1,max = diag(#1,max , · · · , #n 1,max )
then every solution of the closed-loop adaptive system is globally then the origin of the adaptive system is globally asymptoti-
bounded for all |#i (t0 )|  #i,max . cally stable for all |#i (t0 )|  #i,max .
A0 = Aol e = (#0 (t) + #0,max )p 1 bol C > .
#0,max pbb1 bol C0> , A(t) bb 0
ẋp (t) = ap xp (t) + bp v(t), |ap | Plant
ā ẋp(t) = aAppxxpp(t) v(t) |ap |  ā
(t)++bbppv(t), ẋp (t) = Ap xp (t) + bp v(t)
" # " #" # " #
ẋ⌘ (t) = A⌘ x⌘ (t) + b⌘ u(t), ẋ⌘ (t) = A⌘ x⌘ (t) + b⌘ u(t), ẋ
ẋ =
=hAAololxx+ bol u
ẋp Ap bp c>⌘ x p b p d ⌘ i + B⇤(t)y + bol r
T
v(t) = c⌘ x⌘ (t) Unmodeled v(t) = c⌘ x⌘ (t) T = + u =) >> >
ẋ⌘ 0 A⌘ x⌘ b⌘ yyi =
= cCi 0x ⌘ Ci x
Dynamics
G⌘ (s) , c>
⌘ (sInxn A⌘ ) 1
b⌘ G⌘ (s) , c>
⌘ (sInxn A⌘ ) 1
b⌘
m (t) Reference
= am xm (t) + bm r(t), amModel
<0 ẋm(t) = aAmmxxmm(t)
(t)++bbmmr(t),
r(t) am < 0 ẋm (t) = Am xm (t) + bm r(t)
>
u(t) = ✓(t)xp (t) + kr r(t)Control Law ✓ (t)x
u(t) = ✓(t)x p (t)
p (t) ++ kr kr(t)
r r(t) u(t) = ✓ > (t)xp (t) + kr r(t)
˙ = Proj(✓(t), xp (t)e(t)), > 0
✓(t) ˙ = MProj(✓(t),
✓(t) 1
w xp (t)e(t)), >0 ˙ = M 1w
✓(t)
Adaptive Law
wi (t) = Proj {M ✓}i , {M x p b>
m P e}i wi (t) = Proj {M ✓}i , {M xp b>
m P e}i
c⌘ > 2 R1⇥m , x⌘ 2 Rm⇥1 , and A⌘ 2 Rnxn , Am 2 R2⇥2 is Hurwitz
>
p
C = [c0 c1 ] , CP 1 C > = I, M = pbb CP 1 with bm ⌘ bp 2 <[n⇥1] , A> m P + P Am < Q, = 0 P, 0 > 0, pbb = b>
m P bm
" # " #
>
Ap bp c⌘ bp d ⌘ h i h i
1
Aol = , bol = , Ci> = c> >
i 0 , B = pbb bol · · · bol , C = [C 0] , ⇤ = diag(#0 , · · · , #n 1 )
0 A⌘ b⌘

H. Hussain,, “Adaptive Control in the presence of Unmodeled Dynamics,” MIT Thesis, June 2017. 13
 American Control Conference 2019

Three Key Elements

y
Proj(✓, y)

⌦A ⇣ 02
⌘
k✓k2 ✓max
r 0 +"2
⌦1 2"✓max

Parameter is
✓
Projection Modification ✓⇤
bounded for all
˙ =
✓(t) Proj (✓(t), xp (t)e(t)) , >0
⌦0 time

0 = 0, ✓ = ✓max

Transformation of States

Transformed Error " # " #

ė = Am e + bp ✓e xp + bp ⌘ > E˙0 (↵00 + #e0 )E0 + (a1 + #e1 )E1 + pbb ⌘
E ⌘ Ce =
E˙1 A 0 E1 + a0 E0 m

Transformed Parameter Pertubation is

#̇i = 0
Proj(#i , (Ei + mi )E0 ) isolated to
✓˙ = M 1 ẇ # ⌘ M✓
crucial scalar
ẇi = Proj {M ✓}i , {M xp b>
m P e}i , i (Ei + mi )E0 state
r=0 P z1
Z1 (s)
+
>
p
C = [c0 c1 ] , CP 1
C > = I, M = pbb CP 1
with bm ⌘ bp 2 <[n⇥1] , A>
m P + P Am < Q, = 0
P, 0
> 0, pbb = b>
m P bm

e1 (t, z1 )
Absolute Stability Framework
e z1> 1 (t, z1 ) 0
e1 (t, z1 )
r=0 P z1
+
Z1 (s) z1> e1 (t, z1 ) Perform loop transformations, and
enforce a stable feedback
interconnection between linear
subsystem and nonlinearity isolated to
z1

e1 (t, z1 ) feedback path

z1> e1 (t, z1 ) 0
e1 (t, z1 )
H. Hussain,, “Adaptive Control in the presence of Unmodeled Dynamics,” MIT Thesis, June 2017. 14
z1> e1 (t, z1 )
 Reference Model

Gm (s) = C(sI
1
Am ) bm American
m(t) Control Conference 2019
Gol (s)

Robust Adaptive Control Problem Unmodeled Dynamics Plant

ChapterE (t)7. Summary
Reformulated
P > 1 1
v(t) 1
y(t) + P
G⌘ (s) = c⌘ (sI A⌘ ) b⌘ pbb Gp (s) = C(sI Ap ) bp
r(t) = 0 +

#>(t) Reference Model

Adaptive Gain 1
m(t)
Gm (s) = C(sI Am ) bm
#̇i = i (t)E0 yi , |#i (t)|  #i,max

Gol (s) with nth -order plant whose states are acces-
Figure 7-1: Robust adaptive control problem
sible, in the presence of unmodeled dynamics.
Unmodeled Dynamics Summ:fig:ControlProb_vecT
Plant
P 1 v(t) 1
y(t) + P E (t)
G⌘ (s) = c>⌘ (sI A⌘ ) b⌘ pbb1 Gp (s) = C(sI Ap ) bp
Summ:eq:1
r(t) = 0 + of the state errors with (1) |#i (t)|  #i,max . Second, a nonsingular transformation is
employed that allows a tractable relation between parameter errors and state errors.
Namely,
#>(t)
Summ:eq:2a
ẏ1 = Am0 y1 + a0 y0 (2a)
Adaptive Gain Summ:eq:2b
#̇0 = 0 E0 y0 , 0  0 (t)  . 0
(2b)
#̇i = i (t)E0 yi , |#i (t)|  #i,max

Figure 7-1: Third,

Robustthe adaptive
underlying closed-loop adaptive with
control problem systemnisthrelated
-ordertoplant
an absolute
whosestability
states are acces-
framework, which in turn enables the application of the well–known circle criterion
sible, in the presence
to result inofglobal
unmodeled
asymptoticdynamics. Summ:fig:ControlProb_vecT
stability. The adaptive system in Figure 6-1 can be
rewritten as (see Figure 7-1 and [56, §6.2] for details)

e ẋ = (A0 + B(t))x + A(t)x e

ẋ = A0 x + B(t)y(t) ẋ = A0 + A(t) x B1 1(t, y1)
A0 Hurwitz Summ:eq:1 A0 Hurwitz
of thekB(t)k
state errors
 B̄
with
e
kA(t)k > ✏ (1) |# (t)|y 
i 1 =#
>
Ci,max
1x . Second, a nonsingular
e
kA(t)k ✏
transformation
Summ:eq:cl1 is
(7.2)B(t)x 2 [ ⇤max , ⇤max ]
employed y(t) 2that
L2 allows a tractable 1 (t, yrelation
1) = between
⇤1(t)y 1 parameter errors and state errors. e
lim kA(t)k =0
t!1
Namely,
1
G1(s) , C1(sI
>
Adaptive Laws & Input- A0 ) B 1 Use Absolute Stability
Output Stability Framework
where the forward path contains a stable,
Closed-Loop linear, and time-invariant
Adaptive System dynamics, andSumm:eq:2a
a time-varying nonlinearity
0
ẏ1 = Acomprised
m y1 + aof 0 ythe
0 adaptive parameter is isolated to the (2a)
feedback path. Furthermore, the nonlinearity is shown to satisfy a sector constraint. Summ:eq:2b
That is, 1(t, y1) = #̇0⇤1=
(t)y1 20[E0⇤y1,max  ]. 0 (t)  0 .
0 , , ⇤01,max (2b)
Thus, except for the presence of the time-varying matrix A(t),e we note that the 15

Third, the underlying closed-loop

adaptive system adaptive
in (7.2) is identical to the system is related
nonlinear system to an
considered absolute
in the devel- stability
opment of the Circle Criterion, the first absolute stability criterion that extended the
framework, which in turn enables the application of the well–known circle criterion
 American Control Conference 2019

How do I actually apply this?

Second—Order Plant

Theorem 1(b). Suppose r(t) = 0 8t. Consider the adaptive system described by
ẋ = (A0 + Ã(t))x + B1 ⇤1 (t)y1
y1 = C1> x .
G1 (s) , C1> (sI
1
A0 ) B1
If the transfer function matrix
Z1 (s) , (I + ⇤1,max G1 (s))(I ⇤1,max G1 (s))
1

is strictly positive real with ⇤1,max = diag(#1,max , · · · , #n 1,max ), then the origin of the adaptive system is globally
asymptotically stable for all |#i (t0 )|  #i,max .

▫ Steps to design a robust adaptive controller

1 Using the reference model, compute pbb1 and transformation matrix C
2 Assemble
2 closed-loop dynamics
3 and derive G1 (s)
A⌘ b⌘ pbb1 #0 c>
0 b⌘ pbb
1
6 7
6 bp c > Ap 0 7 =) G1 (s)
4 ⌘ 5
0 c>
1 0
#0 = #0,max

Let Z1 (s) , (I + #1,max G1 (s))(I #1,max G1 (s))

1
3
4 Find conditions on (A⌘ , b⌘ , c>
⌘ ) or parameter bounds #i,max such that Z1 (s) is SPR for all admissible plant
parameters.
Easy to satisfy & check
If G1 (s) is Hurwitz, ↵  #1 (t)  , with ↵ < 0 < and the Nyquist plot of G1 (s) lies in the interior of the
disk D(↵, ) then Z1 (s) is SPR and Theorem 1(b) holds.

>
p
C = [c0 c1 ] , CP 1
C > = I, M = pbb CP 1
with bm ⌘ bp 2 <[n⇥1] , A>
m P + P Am < Q, = 0
P, 0
> 0, pbb = b>
m P bm

H. Hussain,, “Adaptive Control in the presence of Unmodeled Dynamics,” MIT Thesis, June 2017. 16
 s2 + ⌧6 s + 12
⌧2
It follows from the main result of this chapter, that a time-delay margin can be derived
for this general adaptive system simply applying the circle criterion to the transfer
function G1(s) in (6.22). It can be shown from (6.55), (6.56) and (6.63) that the
American Control Conference 2019
closed–loop adaptive system in (6.22) is defined completely byb
(⇣p , !p , kp ) = (1, 0.133, 0.16)

Numerical Example
2 3 (⇣m , !m , km ) = (1, 0.4, 0.16)
0 1 0 0 0
(#0,max , #1,max ) = (6, 1.4)
6 12 2#0,max !m #0,max 3/2 7
Time Delay
6 2!m
p
6 ⌧2 2 7
6 ⌧ km (!m +1) km km ! m 2 +1 7
6 7
6 0 0 0 1 0 7 (6.64)
6 7
▫ Demonstrate applicability 6 using ⌧ numerical example 2⇣p !p from #0,max previous 7 work (⌧mm = 4ms )
6 12km 2#0,max !m 2!m
3/2
7 ?
0 2 +1
!m
!p 2 p 2 +1
!m
4
[MM] H. S. Hussain, Y. Yildiz, M. Matsutani, A. M. Annaswamy, and E. Lavretsky, qadaptive
“Computable delay margins for 5 vol. PP, no. 99, pp. 1–1, 2017.
systems with state variables accessible,” IEEE Transactions on Automatic Control,
4 +6! 2 +1
!m
▫2 Order Plant in the presence of an 2 input time delay –Use 2 Order Padé approximation
1
nd 0 0 0 0 nd
m
!m (!m 2 +1)

▫ Plant contains parametric uncertainty

with a corresponding and
transfer can be stable or unstable
function
Chapter
✓ p nth –order
6.Chapter th –order
◆ 6. nRobust Adaptive
Robust Control
Adaptive Control
!m 4 +6! 2 +1
!m 6s
2 +1
!m
m
s2 ⌧
+ ⌧122
G1(s) = ⇣ ⌘ . (6.65)
6s 12 2!m 6s 12
ℑ s + !p2 s2 +
s + 2⇣p !pℑ ⌧
+ ⌧2
+ s+ ! 2 +1
2
sℑ ⌧
+ ⌧2
#0,max
ℑ
m

1.0 = 1.0 1.0 1.0 =

Second-order Unmodeled Dynamics Let us consider the plant in (4.1) and
D(presence
(6.55) in the 1.4, 1.4)
of second-order unmodeled dynamics, described by

0.5 0.5 !n2 0.5 0.5

G⌘ (s) = (6.66)
s2 + 2⇣⌘ !n s + !n2

1 and1 the reference model1 in (4.2) and (6.56). 1 1 In order to

1 determine the robust1 class 1
- - - -
1.4 , !n ) of unmodeled dynamics
(⇣⌘1.4 1.4 for which
1.4 1.4the condition1.4
in Theorem 9 is satisfied,
1.4 we ℜ 1.4
ℜ
-1 -1 1 -1 1 -1 1 1
must compute G1(s) in (6.22). These closed–loop dynamics can be computed directly
since (c>⌘ , A⌘ , b⌘ ) are given by (6.66) and bounds on the parameter uncertainty in the
plant are known. Thus, the stability of the overall adaptive system can be determined
- 0.5 - 0.5 - 0.5 - 0.5
b
⇥ A0B1
⇤
dynamics are written in the form
τ C1 D τ τ τ
0.1349 0.1349 0.1269 0.1269
⇣p,actual = 1 0.2268 0.2268 0.2134 0.2134
0.242 70
0.242
⌧ ? = 245ms 0.2445 0.2445 ⌧ ? = 230ms 0.23 0.23
ζp = 1 ζp =- 1.0
1 - 1.0
0.2449 ζ = −1
p 0.2449
1.0
ζp -= −1 - 1.0
0.2304 0.2304

(a) Stable plant,

(a) Stable
#0,max plant, (b) Unstable (b)
plant,
Unstable
#0,max plant,
Sufficient frequency domain criterion can be
= 6 #0,max = 6
checked graphically for lower order
= 6 #0,max = 6

systems à Delay margin can be derived using Nyquist Plot & Circle Criterion as
Figure 6-3:Figure 6-3:
Nyquist plotNyquist
of G1(j!)
plot &ofCircle
G1(j!)Criterion
& CircleD(Criterion
1.4, 1.4)Dfor time–delay
( 1.4, 1.4) for in
time–delay in
⌧¯? = min (max ({⌧ | G1 (j!, ⇣p , ⌧ ) 2 D( #1,max , #1,max )}))
§6.4.2. §6.4.2. ⇣p

H. Hussain,, “Adaptive Control in the presence of Unmodeled Dynamics,” MIT Thesis, June 2017. 17

SimulationSimulation
Studies Studies
 American Control Conference 2019

Simulation Studies
⇣p,actual = 1
⌧ ? = 245ms
Time Delay

▫ Improved analytical delay margin from ⌧mm ?

= 4ms to ⌧ ? = 245ms
▫ Simulation studies validate theoretical derivations from Time Delay numerical example

⌧ < ⌧ ? =) Asymptotic Stability ⌧ ⌧ ? =) Instability

Methodology & approach extends to the case of real (not approximated) time—delay
and provides a practical and analytically computable delay margin

H. Hussain,, “Adaptive Control in the presence of Unmodeled Dynamics,” MIT Thesis, June 2017. 18
 American Control Conference 2019

Why Adapt?
Parameter Convergence
9
Reference Model ẋm (t) = Am xm (t) + br(t) >
>
>
>
=
Plant ẋp (t) = Ap xp (t) + bu(t)
Goal: xp (t) ! xm (t) (= ✓(t) ⌘ ✓?
Control Input u(t) = ✓(t)> xp (t) + r(t) >
>
> Ap + b✓ ? = Am
>
; | {z }
Closed-loop ẋp = (Ap + b✓> (t))xp + br(t)
matching condition (1)

Theorem (Parameter Convergence [Kokotovic et al., 1985]). Consider the system (✓(t) ! ✓ ? )
" # " #" #
>
ė A b!(t) e
= .
˙ µ!(t)h> 0
with , ✓ ✓? and W m (s) , h> (sI A) b. Let !(t) be bounded, almost periodic, and persistently exciting.
1

Then there exists a µ? > 0 such that for all µ 2 (0, µ? ], the origin of the system is exponentially stable if
" Z !#
T
min < i !(t)W m (s)!(t)> d⌧ > 0.
i 0

   
ẋp Ap bc> xp 0
+ Unmodeled Dynamics: = ⌘ + r(t) @✓ ? s.t. (1) is satisfied
ẋ⌘ b⌘ ✓ ?> A⌘ x⌘ b⌘

Condition 1. If 9✓ such that the equality

> 1
I Gp (s)G⌘ (s)✓ Gp (s)G⌘ (s)r(t) = Gm (s)r(t)
is satisfied, then an operator matching condition is said to exist for ✓. Furthermore, if the adaptive gain ✓(t) = ✓,
then the tracking error e = xp xm is equal to zero.

Derived notionally equivalent matching condition, in frequency domain, for which

local convergence Theorem holds & original control goal of tracking is achieved.

H. Hussain,, “Adaptive Control in the presence of Unmodeled Dynamics,” MIT Thesis, June 2017. 19
 American Control Conference 2019
(⇣p , !p , kp ) = (1, 1, 1)

Simulation Studies (⇣m , !m , km ) = (1, 3, 1)

Parameter Convergence (Rohrs’ Unmodeled

Chapter 8.6.7. Dynamics)
Why
Why Adapt?
Adapt?

10 r(t) = 100 sin(!t), ! 2 { ⇡4 , ⇡2 , 3⇡

2 , ⇡} 8 r(t) = 100 with 0
2 [0.06, 0.3]
8 6
6
|E0 (t)| |E0 (t)| 4
4
2 2
Tracking errors exponentially decay
0 0

-0.3 -0.3

ϑ(t) -2.8 ϑ(t) -2.8

-6.8 -6.8
1000 10 20 30 40 50 2000 5 10 15 20
Adaptive Parameters find a steady state
50 150
“matching condition” even with model mismatch
r(t) 0 r(t) 100
-50 50

-100 0
0 10 20 30 40 50 0 5 10 15 20
time time
-0.3 -0.3

ϑ1 (t) ϑ1 (t)

-2.8 -2.8
-6.8 -0.9 -6.8 -0.9
DC ϑ0 (t) DC ϑ0 (t)

2 , ⇡} s Figure
and } 6-13:
= 600 Simulation with r(t) = 100 and ˝ = 300 ◊ {2, 4, 6, 8, 10} with “ Õ = .
˝
Figure 8-3: Simulation with r(t) = 100 sin(!t) with ! 2 { ⇡4 , ⇡2 , 3⇡ rad
} Adaptive Controller is able to asymptotically track even in the presence of both
r̄ 2
with = 1002 .
0
parametric and nonparametric uncertainties, due to the existence of an operator
matching
Fig. condition
0
0
0
1 #0&(0) parameter convergence
#1 (0) m0 (0) m1 (0) E0 (0) of Eits –ı clearly
kind and
1 (0) demonstrates
allows application the such
of well-known results benefit of of averaging.
as the method
adaptation
8-5(a) over
0.2952 non-adaptive
0.04919 -6.8 -2.8 0 control
0 designs
0.0824Thus,
-0.1201when ◊ exists and the input is persistently exciting, we can analytically assert local
Figure 8-3–6-8(b) 0.2952 0.04919 -3.4 -1.4 82.32 120 0.0824asymptotic
-0.1201 stability of the origin.
For those particular situations, it is easy to show that the robust adaptive controller
proposed herein yields superior performance when compared to a non-adaptive robust
H. Hussain,, “Adaptive Control in the presence of Unmodeled Dynamics,” MIT Thesis, June 2017. 20 con-
Table 8.1: Initial conditions for simulation studies in Figuretroller,
8-5 i.e. ◊(t) vs. ◊ . This follows immediately from the fact that the main motivation
lqr
for implementing an adaptive controller stems from the plant having parametric uncer-
tainty. Therefore, when ◊lqr is designed from a nominal (Ap , bp ) instead of the actual plant,
 American Control Conference 2019

Summary
Robust Adaptive Control

▫ Solved an open problem:

– Rigorously proved global boundedness of a closed-loop adaptive system comprised of a LTI nth–order plant, whose
state variables are accessible, in the presence of unmodeled dynamics
– Class of unmodeled dynamics for which the system is robust to is shown to be analytically computable.

▫ Reformulated the robust adaptive control problem into a well-known stability

framework
– Employed the Circle Criterion to analyze stability of the solutions and proved global boundedness
▫ Sufficient frequency domain criterion guarantees global boundedness
▫ Verified fundamental trade-off between adaptation & robustness
▫ Validated analytical results via simulation
– Demonstrated applicability and practicality of the result
▫ Extends to Multiple-Input systems and systems with time-delay
▫ Proved that tracking is still possible even in the presence of unmodeled
dynamics, due to the existence of a novel operator matching condition

▫ Limitations & Future Work:

– Currently applies to LTI plants whose states are accessible, apply methodology to output feedback
– Consider other classes of non-parameteric uncertainties and/or nonlinear plants

We have showed the extension of this result to such systems at Boeing.

21
 American Control Conference 2019

Fundamental Trade-off
Scalar Numerical Example (Roll Dynamics)

▫ Adaptive system is robust to all unmodeled dynamics shown in shaded regions ( for fast
roll dynamics and for slower) given adaptation bound ✓max
bm pm
s am

Time Aircraft Roll Dynamics P e

Delay
P acmd a p +
pcmd +
+
Analysis Parameter (⇣, !n ) a

✓
Design Parameter
|✓(t)|  ✓max

þ Linear Robustness Margins

(>0.6, ⇠50)
þ Not Conservative
(>0.6, ⇠75)
þ Validated in Simulation
!⌘

Larger adaptation bound è smaller robust class of unmodeled dynamics

23