Indian Institute of Technology Bombay

Department of Aerospace Engineering

AE 666 Adaptive and Learning Control Systems
Spring 2024-2025

Tutorial 3 Solution March 25, 2025

1. Consider the following scalar constant gain system

y = θ∗ u + d,

where d is a bounded unknown disturbance and u ∈ L∞ . The adaptive law for estimating θ∗
derived for d = 0 is given by
θ̇ = γϵu, ϵ = y − θu,
where γ > 0. How does the presence of a non-zero bounded unknown disturbance d affect the
stability of the parameter estimate θ (t), and what phenomenon results from this instability?

Solution: The parameter error equation,

˙
θ̃ = θ̇ − θ̇∗
= γϵu
= γ (θ∗ u + d − θu) u
= −γu2 θ̃ + γdu.

Let us consider a bounded input u of the following form,

1
u= √ .
1+t
For γ = 1 and θ∗ = 1, consider following d,
1 1
d= 1/4
−√ ,
(1 + t) 1+t

such that d → 0 as t → ∞. We obtain,

1
y= , y → 0 as t → ∞,
(1 + t)1/4

˙ 1 1 1
θ̃ = − θ̃ + 3/4
− ,
1+t (1 + t) 1+t
˙
(1 + t) θ̃ = −θ̃ + (1 + t)1/4 − 1,
d n o
(1 + t) θ̃ = (1 + t)1/4 − 1,
dt
integrating both sides,
4
(1 + t) θ̃ = (1 + t)5/4 − t + C,
5
AE 666 Tutorial 3 Solution March 25, 2025

4 t C
θ̃ = (1 + t)1/4 − + , θ̃ → ∞ as t → ∞.
5 1+t 1+t

2. Check whether the given ϕ (t)’s are persistently exciting/ exciting over a finite interval/ not
exciting.
T
(a) ϕ (t) = 1 e−t .


 T
(b) ϕ (t) = sin t cos t .

Solution:
(a) We evaluate the following integral,
e−t − e−(t+T )
" #
Z t+T 
1 e−τ
 T
dτ = 1 −2t 1 −2(t+T ) .
t e−τ e−2τ e−t − e−(t+T ) e − e
2 2
To satisfy the following inequality,
Z t+T 
e−τ
  
1 1 0
dτ − α > 0,
t e−τ e−2τ 0 1
the choice of α ∈ R+ is dependent on time t, therefore the given ϕ is exciting over a finite interval.
(b) We evaluate the following integral,
Z t+T 
sin2 τ

sin τ cos τ
dτ =
t sin τ cos τ cos2 τ
 
1 1
1 T − sin (t + T ) cos (t + T ) + sin t cos t − cos 2 (t + T ) + cos 2t
2 2 .

2
 1 1
− cos 2 (t + T ) + cos 2t T + T + sin (t + T ) cos (t + T ) − sin t cos t
2 2
Choosing T = 2π, we obtain,
Z t+2π 
sin2 τ
  
sin τ cos τ π 0
dτ =
t sin τ cos τ cos2 τ 0 π
To satisfy the following inequality,
Z t+2π 
sin2 τ
  
sin τ cos τ 1 0
dτ − α > 0,
t sin τ cos τ cos2 τ 0 1
the choice of α < π is independent of time t, therefore the given ϕ is persistently exciting.

3. Consider the second-order differential equation

ẋ = −uuT x + f (x) ,
where T
T
f (x) = x2 x21 + x22 −x1 x21 + x22
 

x = x1 x2 , ,
and u : R+ → R2 . Is u satisfying the persistence of excitation condition necessary and sufficient
for uniform asymptotic stability of the origin of the given nonlinear system? Explain.

2
AE 666 Tutorial 3 Solution March 25, 2025

Solution: Is u satisfying the persistence of excitation (PE) condition necessary? We choose a non-PE
 T
u, u = 0 1 , and substituting it in the given nonlinear system, we obtain

x2 x21 + x22

      

ẋ1 0   x1
=− 0 1 + ,
ẋ2 1 x2 −x1 x21 + x22

ẋ1 = x2 x21 + x22 ,

ẋ2 = −x2 − x1 x21 + x22 .

Investigating the stability of the origin by choosing the following Lyapunov function candidate,
1 1
V = x21 + x22 ,
2 2
the time-derivative along the trajectories,

V̇ = −x22 .
n o
Let E = (x10 , 0) : V̇ = 0 be the largest invariant set.

ẋ1 = 0 =⇒ x2 = 0,
ẋ2 = 0 =⇒ x1 = 0.
n o
Therefore, E = (0, 0) : V̇ = 0 is the largest invariant set; hence, the origin is uniformly asymptotically
stable. This implies that u satisfying the PE condition is not a necessary condition.
Is u satisfying the persistence of excitation (PE) condition sufficient? We choose a PE u, u =
 T
sin t cos t , and substituting it in the given nonlinear system, we obtain,

ẋ1 = −x1 sin2 t − x2 sin t cos t + x2 x21 + x22 ,

ẋ2 = −x1 sin t cos t − x2 cos2 t − x1 x21 + x22 ,

let the solution of x1 and x2 in polar form be x1 = r cos t and x2 = −r sin t, (r > 0),

x1 ẋ1 + x2 ẋ2 = 0,
x21 + x22 =C, C > 0,
 T
hence, x1 = r cos t and x2 = −r sin t is the solution of the given nonlinear system when u sin t cos t
is PE. Therefore, u satisfying the persistence of excitation (PE) condition is not a sufficient condition.

4. Consider a system with the first-order differential equation

ẋ = ax + bu,

where x ∈ R is the system state and u ∈ R is the control input of the system. The system
parameters a and b are unknown, however, sgn (b) is known. Let the control law be

u = kx x + kr r,

3
AE 666 Tutorial 3 Solution March 25, 2025

drives the system state x to follow a reference state xr that is driven by the following first-order
differential equation,
ẋr = ar xr + br r,
where r is a bounded piece-wise continuous signal, ar < 0 and br ∈ R. Let there exists real
constants kx∗ and kr∗ satisfying, ar = a + bkx∗ and br = bkr∗ . Further let the state-tracking error be
denoted by e is defined as e = x − xr , and two prediction errors ϵa = ar − â − b̂kx and ϵb = br − b̂kr ,
where â and b̂ are the estimates of a and b, respectively. Design update laws for kx and kr such
 T
that the origin of the error vector e ϵa ϵb is asymptotically stable.

Solution: The tracking error dynamics,

ė = ar e + bk̃x x + bk̃r r,

where k̃x = kx − kx∗ and k̃r = kr − kr∗ . Let ã = â − a and b̃ = b̂ − b, defining Lyapunov function
candidate,
1 |b| |b| 1 1
V = e2 + k̃x2 + k̃r2 + ã2 + b̃2 ,
2 2 2 2 2
the time-derivative along the error trajectory,
˙
V̇ = ar e2 + bk̃x xe + bk̃r re + |b| k̃x k̇x + |b| k̃r k̇r + ãâ˙ + b̃b̂,

simplifying the ϵa and ϵb as below,

 
ϵa = a + bkx∗ − â − b̃ + b kx ,
= −ã − bk̃x − b̃kx ,
 
ϵb = bkr∗ − b̃ + b kr ,
= −bk̃r − b̃kr ,
choosing the adaptation laws as below,

k̇x = −sgn (b) xe + sgn (b) ϵa ,

k̇r = −sgn (b) re + sgn (b) ϵb ,
â˙ = ϵa ,
˙
b̂ = kx ϵa + kr ϵb ,

putting in V̇ , we obtain,

V̇ = ar e2 + bk̃x ϵa + bk̃r ϵb + ãϵa + b̃kx ϵa + b̃kr ϵb ,

   
= ar e2 + ϵa bk̃x + ã + b̃kx + ϵb bk̃r + b̃kr ,
= ar e2 − ϵ2a − ϵ2b .

V > 0 and V̇ ≤ 0 implies that e, k̃x , k̃r , ã, and b̃ are bounded and e, ϵa , and ϵb are square integrable.
˙ and b̂˙ are bounded, this further implies
Further, x, ϵa , and ϵb are bounded, this implies that ė, k̇x , k̇r , â,
 T
that ϵ̇a and ϵ˙b are bounded. Therefore, e ϵa ϵb ≡ 0 is asymptotically stable.

4
AE 666 Tutorial 3 Solution March 25, 2025

5. Consider the following algebraic model of a system,

y (t) = ϕ (t) w∗ ,

where y ∈ R is the measurable output of the given system, ϕ ∈ R2 is a known vector, and
w∗ ∈ R2 is an unknown constant vector.
(a) Design a gradient-based adaptation law to update the estimate of the unknown vector w∗ .
Verify the convergence of the parameter estimates to their ideal values using numerical
simulations taking ϕ (t)’s given in Q.2.
(b) Modify the adaptation law designed in (a) using the least squares method and compare the
evolution of parameter estimates using numerical simulation.

Solution:
1. Gradient-based adaptation law,
ẇ = γϕ y − ϕT w

0 2

1.5
-1

-2
0.5

-3 0
0 2 4 6 8 10 0 2 4 6 8 10

T
Figure 1: Tracking error and parameter estimation for ϕ (t) = 1 e−t , γ = 1.


0 2.5

-0.2 2

-0.4 1.5

-0.6 1

-0.8 0.5

-1 0
0 5 10 15 0 5 10 15

 T
Figure 2: Tracking error and parameter estimation for ϕ (t) = sin t cos t , γ = 2.

2. Least-squares based adaptation law,

ẇ = γRϕ y − ϕT w , Ṙ = −RϕϕT R

5
AE 666 Tutorial 3 Solution March 25, 2025

0 2

1.5
-1

-2
0.5

-3 0
0 5 10 15 0 5 10 15

T
Figure 3: Tracking error and parameter estimation for ϕ (t) = 1 e−t , γ = 2, R0 = I2 .


0 2.5

-0.2 2

-0.4 1.5

-0.6 1

-0.8 0.5

-1 0
0 5 10 15 0 5 10 15

 T
Figure 4: Tracking error and parameter estimation for ϕ (t) = sin t cos t , γ = 5, R0 = I2 .

6. Consider a system with the following plant dynamics, incorporating actuator saturation
   
0 1 0
ẋ = x+ R (u) ,
0 0 1 s

where the actuator saturation function Rs (u) is defined as,


u,
 if 0 < u < 5
Rs (u) = 5, if u > 5 .

0, if u < 0


Design an MRAC by choosing a suitable reference model, such that the origin of the state tracking
error is asymptotically stable.

Solution: Let the control deficiency be defined as,

∆u = Rs (u) − u,

6
AE 666 Tutorial 3 Solution March 25, 2025

the modified plant dynamics is,

    
0 1 0 0
ẋ = x+ u+ ∆u,
0 0 1 1

let the reference model be,  

0
ẋr = Ar x + br r + ∆u,
1
the tracking error dynamics, driven by the control law u = kTx x + kr r,

ė = Ar e,

where system parameters are known hence, the controller gains kx and kr are known. Since, Ar is
Hurwitz, the origin of the tracking error e is asymptotically stable.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . End . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .