AIAA 2015-0608

AIAA SciTech
5-9 January 2015, Kissimmee, Florida
AIAA Guidance, Navigation, and Control Conference

Application of a Novel Scalability Notion in Adaptive

Control to Various Adaptive Control Frameworks

Simon P. Schatz, Tansel Yucelen, Benjamin Gruenwald, and Florian Holzapfel

In adaptive control theory it is a well-known phenomena that nonidentical command

profiles entail nonidentical closed-loop responses of these adaptive systems. While adaptive
controllers provide a viable methodology to control uncertain dynamical systems, this lack
Downloaded by CARLETON UNIVERSITY LIBRARY on July 31, 2015 | http://arc.aiaa.org | DOI: 10.2514/6.2015-0608

of predictability is a significant disadvantage, in particular in terms of certification of such

control methods. Consequently, achieving predictable closed-loop responses of adaptively-
controlled systems is of grand practical interest. For this purpose, we recently introduced a
method1 to scale the learning rates of the adaptive weight update laws in order to achieve
predictable closed-loop performances for nonidentical, but scalable command profiles. This
paper applies the proposed methodology to a model of the longitudinal motion of a Boeing
747 aircraft and simulations for diverse adaptive control schemes illustrate the efficiacy
of the proposed scalability notion, which may be a further step towards validation and
verification of these adaptive control frameworks.

I. Introduction
In this paper, direct model reference adaptive control (MRAC)2, 3 is considered. Adaptive controllers
require less modeling information in comparison to fixed-gain controllers as the controller gains are tuned
online driven by the tracking error between the systems output (respectively, state) and the reference models
output (respectively, state). Hence, controllers employing adaptive control laws have the ability to deal with
uncertainties e.g. resulting of unknown nonlinearities or imprecisely modeled system parameters.
Although adaptive controllers show good results for well-tuned cases, it is well-known in the adaptive
control community that adaptive controllers are sensitive to the level of excitation, particularly in the tran-
sient phase when the adaptive control method learns the uncertainty. On the other hand, employing high
learning rates for the adaptive weight update laws in order to increase the adaptations speed in the transient
phase may result in unacceptable control input signals due to high-frequency content in the control channel.4
Nowadays, a lot of research is conducted towards these problems and modifications such as pseudo control
hedging,5 low-frequency learning adaptive control,6 and L1 adaptive control7 have been introduced in order
to employ high learning rates. Frameworks to achieve improvement of the transient performance of adaptive
controllers were further introduced, namely among others closed-loop reference models8, 9 and the command
governor adaptive control framework.10
In addition, high excitation of the regressor vectors in the adaptive weight update law have a similar effect
as high learning rates, which results of the nature of the adaptive weight update laws. As a consequence,
tuning the adaptive controller for good transient behavior over the whole envelope of permissible system states
is a challenging task. Although some adaptive control architectures show a certain level of predictability and
are approximately scalable, for example L1 adaptive control7 and the command governor framework,10 the
schemes introduced above are all related to improving the performance of adaptive controllers, particularly
This research was supported in part by the University of Missouri Research Board.
Graduate Research Assistant, Institute of Flight System Dynamics, Technische Universit at Munchen, 85748 Garching,
Germany, simon.p.schatz@tum.de
Assistant Professor, Missouri University of Science and Technology, Department of Mechanical and Aerospace Engineering,

Rolla, MO 65409, USA, tyucelen@mst.edu

Graduate Research Assistant, Missouri University of Science and Technology, Department of Mechanical and Aerospace

Engineering, Rolla, MO 65409, USA, bg8t2@mst.edu

Professor, Institute of Flight System Dynamics, Technische Universit at Munchen, 85748 Garching, Germany, flo-
rian.holzapfel@tum.de

1 of 17

American Institute of Aeronautics and Astronautics

Copyright 2015 by Simon P. Schatz, Tansel Yucelen, Benjamin Gruenwald, Florian Holzapfel. Published by the American Institute of Aeronautics and Astronautics, Inc., with permission.
 in the transient phase, but do not directly address predictability of the system response over the envelope
of permissible commands.
We recently introduced a scalability notion in Ref. 1, which shows scalable performance with respect
to nonidentical, but scalable command profiles for a class of direct model reference adaptive controllers
with uncertainties parameterized using linear regressor vectors. This scalability concept relies on scaling the
learning rates relating to the command coefficient of these command profiles. The contribution of this paper
is to execute an illustrative simulation study using the previous theoretical results1 and a model of the lon-
gitudinal motion of a Boeing 747 aircraft. This simulation study incorporates adaptive control architectures
using reference model modifications,8, 9 adaptive controllers with e-modification,12 low-frequency learning
adaptive control,6 and the command governor adaptive control framework.10 This simulation study analyzes
the numerical results in detail and shows the efficiacy of the proposed scalability concept. In comparison to
previous work, this command governor example is evaluated in more detail, emphasizing the approximatly
scalable nature of this framework. From the theoretical point of view, the scalability notion is further shown
for an adaptive control framework employing artificial basis functions.13 This method is also illustrated in
simulations and the effect of the scalability notion is analyzed.
Downloaded by CARLETON UNIVERSITY LIBRARY on July 31, 2015 | http://arc.aiaa.org | DOI: 10.2514/6.2015-0608

II. Scalability for Model Reference Adaptive Control

This section overviews our previous results on scalability in Ref. 1 and hence, the problem formulation
utilized for classical MRAC schemes is presented and the scalability notion for MRAC is introduced. Ad-
ditionally, the application of the scalability notion on different MRAC schemes, specifically -modification
and e-modification adaptive control architectures,11, 12 adaptive control architectures with low-frequency
learning,6 adaptive control architectures employing closed-loop reference models,8, 9 and command governor-
based adaptive controllers,10 is overviewed shortly in order to elucidate the previous results before using
these adaptive control frameworks for illustrative simulations displayed in Section III. Finally, an adaptive
control scheme using artificial basis functions13 is shown to be scalable in the sense presented in this paper.
The nomenclature used for this paper is fairly standard and according to the nomenclature used in previous
publications, e.g. in Ref. 10.

A. MRAC Problem Formulation

Consider the uncertain dynamical system given by

x(t)
= Ax(t) + Bu(t) + B(x(t)), x(0) = x0 , (1)

where x(t) <n is the accessible state vector, u(t) <m is the control input vector, (x(t)) : <n <m is
an uncertainty, A <nn is a known system matrix, <mm + is an unknown control effectiveness matrix,
and B <nm is a known control input matrix.
It is assumed that the pair (A, B) is controllable and the uncertainty is parameterized as
h i
(x(t)) = WxT WcT w (x(t), c(t)), (2)

where Wx <nm denotes an uncertainty in the system matrix, Wc <lm denotes an uncertainty in
the command input matrix, and w <m denotes a constant disturbance. Note that Wx , Wc , and w are
considered time-invariant. Although the formulation of the uncertainty in (2) represents a class of linear
uncertainties, the overall system including the adaptive control scheme is inherently nonlinear.
Additionally, (x(t), c(t)) <n+l+1 is a known regressor vector. given by

x(t)
(x(t), c(t)) = c(t) , (3)

where c(t) <l is the uniformly continuous bounded command and < is a constant.
The uncertain dynamical system specified in equation (1) is desired to track the reference system given
by

x r (t) = Ar xr (t) + Br c(t), xr (0) = xr0 , (4)

2 of 17

American Institute of Aeronautics and Astronautics

 where xr (t) <n is the reference model state vector, Ar <nn is the desired Hurwitz system matrix, and
Br <nl is the command input matrix. The reference system (4) is desired to be tracked by the uncertain
dynamical system (1) using the control law

u(t) = unom (t) uad (t), (5)

where u(t) <m is the control input, uad (t) <m is the adaptive control input and unom (t) <m denotes
the nominal control input given by

unom (t) = Kx x(t) + Kc c(t), (6)

where Kx <mn is the nominal feedback matrix and Kc <ml is the nominal feedforward matrix chosen
such that A BKx = Ar and BKc = Br . Using (2), (5), and (6) in (1), the uncertain dynamical system is
given by

= Ar x(t) + Br c(t) + BW T (x(t), c(t)) Buad (t),

Downloaded by CARLETON UNIVERSITY LIBRARY on July 31, 2015 | http://arc.aiaa.org | DOI: 10.2514/6.2015-0608

x(t)
(7)

where
h iT
W , WxT Kx WcT + Kc w , (8)

and , Im 1 . Now, the adaptive control law is chosen as

 

T (t)(x(t), c(t)),
uad (t) = W (9)
(t) <(n+l+1)m is the adaptive weight matrix satisfying the adaptive weight update law
where W

(t)
W = (x(t), c(t))eT (t)P B, (0) = W
W 0, (10)

which is driven by the tracking error e(t) <n defined as

e(t) , x(t) xr (t), (11)

and P <nn is the positive definite solution of the Lyapunov equation

Q + AT
r P + P Ar = 0, (12)

where Q <nn is a positive definite design matrix. Hence, the uncertain dynamical system (1), (7) can
now be given as

x(t)
T (t)(x(t), c(t)),
= Ar x(t) + Br c(t) BW x(0) = x0 , (13)

where W (t) , W
(t) W <(n+l+1)m is the adaptive weight estimation error. Stability of direct model
reference adaptive controllers as introduced in this section can be shown according to the literature.2, 3

B. Scalability
In this section, the concept of scalability1 is introduced. The objective of the scalability notion is to achieve
scalable performance for nonidentical, but scalable command profiles c(t).
For the idea of scalability we assume that a control engineer has designed a positive definite matrix Q for
the Lyapunov equation (12) and a learning rate 0 yielding appropriate performance of the adaptive control
system for a specified command history c0 (t). Now, the idea of the scalability notion is to scale the learning
rate corresponding to the specified command profile. Hence, for any scaled command profiles c(t) = c0 (t)
with scalar scaling command coefficients 6= 0 given a Lyapunov design matrix Q it is possible to achieve
scaled system responses with respect to this well-tuned set-up by choosing = 0 /2 .
This can be shown in a straight forward manner by transformation1 of the system dynamics (13), the
reference system (4) and the weight update law (10). In order to execute this transformation, the following
definitions are employed:

3 of 17

American Institute of Aeronautics and Astronautics

 z(t) , x(t)/
z0 , x0 /
zr (t) , xr (t)/
(14)
zr0 , xr0 /
ez (t) , e(t)/
=

Hence,

z(t)
z (x(t), c(t)) = (x(t), c(t))/ = c0 (t) . (15)

Downloaded by CARLETON UNIVERSITY LIBRARY on July 31, 2015 | http://arc.aiaa.org | DOI: 10.2514/6.2015-0608

Using (14) and (15) the transformed system dynamics, the transformed reference system dynamics and the
adaptive weight update law are given by

z(t)
= T z (x(t), c(t)),
Ar z(t) + Br c0 (t) BW z(0) = z0 , (16)
zr (t) = Ar z(t) + Br c0 (t), zr (0) = zr0 , (17)
W (t) = 0 z (x(t), c(t))eT
z (t)P B,
(0) = W
W 0. (18)

Note that the equations (16), (17), and (18) hold for any 6= 0. Further, note that the uncertain system
(13,16) and the reference system (4,17) are scalable in the sense that state histories can be given by a nominal
system response scaled by .
Finally, consider the adaptive weight update law for the original case (10) and for the transformed system
(18). Using (14), (15) and the scaled learning rate = 0 /2 , it can be stated

(t) = (x(t), c(t))eT (t)P B = 0 z (x(t), c(t))eT (t)P B.

W (19)
z

Consequently, (19) shows that using = 0 /2 renders invariance of the adaptive weight update law
with respect to the scaling parameter . Thus, the adaptive weight responses are identical given any . This
deviates from a traditional adaptive control architecture since large regressor vectors (x(t), c(t)) have the
same negative effects, namely undesirable oscillations,4 on most adaptive control frameworks as excessively
large learning rates , which is obvious from (10). Hence, for uncertain dynamical systems of the form (13)
scalable performance can be achieved for any 6= 0 and large commands c(t) do not cause high excitation
of the adaptive weights, yielding a certain level of predictability, which is required for a further step towards
validation and verification of adaptive controllers.

C. Scalability of other MRAC schemes

In this section, the application of the scalability notion to extensions of the MRAC framework is presented. In
particular, the application of the scalability concept to the -modification and e-modification adaptive control
architectures11, 12 is shown. Furthermore, adaptive control architectures with low-frequency learning6 , adap-
tive control architectures employing closed-loop reference models,8, 9 and command governor-based adaptive
controllers10 are modified to achieve predictable performances as shown previously. In terms of brevity of
this paper, these modifications are introduced in short and for more detailed information the reader may be
referred to Ref. 1.

1. , and e modification architectures

The -modification11 was introduced with the claim that it prevented the estimated adaptive weights from
becoming unbounded. In order to achieve this increased robustness with respect to e.g. unmodeled dynamics,

4 of 17

American Institute of Aeronautics and Astronautics

 the adaptive weight update law was modified as

(t) = (x(t), c(t))eT (t)P B W

W (t), (0) = W
W 0, (20)

where > 0 is a damping coefficient used to pull the estimated adaptive weights towards the origin. By
introducing the scaling factor for the learning rate = 0 /2 as introduced in Section B, invariance of the
adaptive weight response with respect to the command coefficient can be achieved.1
The standard MRAC adaptive law was further modified by introducing a time-varying damping coefficient
e ke(t)k2 instead of the constant in (20). Hence the effect of the damping in the so-called e-modification12
was proportional to the tracking error of the system, resulting in the adaptive weight update law given by

(t) = (x(t), c(t))eT (t)P B e ke(t)k W

W (t), (0) = W
W 0, (21)
2

where e > 0. By scaling the learning rate = 0 /2 and the damping coefficient e = 0 /, it can be
shown that the adaptive weight response is invariant to the scaling factor .1
Downloaded by CARLETON UNIVERSITY LIBRARY on July 31, 2015 | http://arc.aiaa.org | DOI: 10.2514/6.2015-0608

Since the adaptive weight update laws for these two modifications are invariant to the scaling factor
and both, control law and reference model, remain unchanged, the uncertain dynamical systems response
is scalable as introduced in Section B.

2. Low-Frequency Learning
The low-frequency learning adaptive control architecture employs a gradient based modification term and a
low pass filter6 . It is claimed that the modification term filters high-frequency content out of the adaptive
weight update law, allowing for the controller to be tuned with high learning rates in order to enable robust
and fast adaptation. The adaptive weight update law is given by

(t) = (x(t), c(t))eT (t)P B [W

W (t) W
f (t)], (0) = W
W 0, (22)

f (t) <(n+l+1)m is the low-pass filtered weight estimate of W

where > 0 is a modification gain and W (t),
satisfying

f (t) = f [W
W (t) W
f (t)], f (0) = W
W 0, (23)

where f <(n+l+1)(n+l+1) is a positive definite filter gain matrix such that max (f ) f,max and
f,max > 0 is a design parameter.
Incorporating the scaling factor into the learning rate as = 0 /2 yields once again invariance of the
adapive weights with respect to the command coefficient .1 Therefore, as discussed in the previous section,
it can be concluded that a system employing this adaptive control framework will have predictably scalable
responses.

3. Reference Model Modification

An approach to improve the transient performance of MRAC controllers by feeding back the tracking error
into the reference model is the so-called reference model modification8, 9 . As a consequence, the uncertain
dynamical system (13) and the adaptive weight update law (10) remain unchanged and are scalable as
presented in Section B. However, the reference model is given by

x r (t) = Ar xr (t) + Br c(t) + Le(t), xr (0) = xr0 , (24)

where L <nn is a positive definite matrix. Employing the relations (14) and scaling the command with
the command coefficient c(t) = c0 (t) scalability for adaptive control architectures with modified reference
models is obtained.1

4. Command Governor Adaptive Control

Finally, the scalability notion is applied to the command governor framework for adaptive control10 . Basi-
cally, an additional command was introduced in order to cancel uncertainties in the transient phase.

5 of 17

American Institute of Aeronautics and Astronautics

 In contrast to traditional adaptive control frameworks, the command is given by
c(t) , cD (t) + cg (t), (25)
where cD (t) <m is the bounded, desired tracking command (the original c(t) from the sections above).
1 T
The additional command cg (t) , Kc1 B T B

B g(t) <m , det(Kc ) 6= 0 is based on a linear system,
which is defined as

(t) = (t) + e(t), (0) = 0, (26)
g(t) = (t) + [Ar In ] e(t), (27)
where (t) <n denotes the command governor states, g(t) <n is the command governor output, and
> 0 is the command governor gain. Since the additional command is applied on both reference model and
nominal controller, the error dynamics of the system do not change and therefore, we have
e(t) T (t)(x(t), c(t)),
= Ar e(t) BW e(0) = x0 xr0 , (28)
Downloaded by CARLETON UNIVERSITY LIBRARY on July 31, 2015 | http://arc.aiaa.org | DOI: 10.2514/6.2015-0608

which can be written as

T (t)(x(t), c(t)) = [B T B]1 B T {Ar e(t) e(t)}.
W (29)
Applying (25), (26), (27), and (29) onto the uncertain system dynamics (13), we have
1 T
= Ar x(t) + Br cD (t) + B B T B

x(t)
B {(t) e(t) e(t)},
x(0) = x0 . (30)
= 0 for and that the overall system is stable.10
It can be shown that (t) e(t) e(t)
Note that although the reference model is modified, the closed loop uncertain system still tracks the
desired reference model given by
x r,D (t) = Ar xr,D (t) + Br cD (t), xr,D (0) = xr0 (31)
as the last term of (30) is appoximately zero for large .
The interested reader may be referred to previous work10 for additional information about the command
governor. Using cD (t) = c0 (t), and = 0 /2 it can be shown that the command governor framework is
also scalable in the sense introduced in this paper.1

D. Application of Scalability Notion using a Framework with Artificial Basis Functions

In this section, the scalability notion is further extended to an adaptive control framework using artificial
basis functions.13, 14 It is claimed that the transient performance is improved by introducing an artificial
regressor vector a (t) <q , which is updated online and is used in an adaptive control law of the form
T (t)(x(t), c(t)) + W
uad (t) = W aT (t)a (t), (32)
where W a (t) <qm is an estimate of the ideal artificial weight Wa <qm , which is an artificial and thus,
nonexisting weight Wa = 0. Then the artificial weight estimation error is given by W a (t) , W
a (t) Wa =

Wa (t). Inserting (32) into the uncertain system given by (7), the system dynamics can be computed as
 
x(t)
= Ar x(t) + Br c(t) B W T (t)(x(t), c(t)) + W
T (t)a (t) , x(0) = x0 . (33)
a

In Ref. 13 it is claimed that the mismatch term W T (t)(x(t), c(t))+ W

aT (t)a (t) is approximately suppressed
by gradient minimization, resulting in the artificial basis function update law given by
aT (t)[B T B]1 B T {e(t)
a (t) = ka W Ar e(t)}, a (0) = a0 , (34)
where ka <+ is a design parameter and a0 6= 0. By applying integration of parts, this equation can be
implemented without requiring differentiation of the system error e(t) or knowledge of its time derivative
e(t)
(see Ref. 14 for more details). The adaptive weight update laws are given by13, 14

(t) = (x(t), c(t))eT (t)P B + (x(t), c(t))aT (t)W

 
W a (t) , W
(0) = W
0, (35)
a (t)
W = a a (t)eT (t)P B, a (0) = W
W a0 , (36)

6 of 17

American Institute of Aeronautics and Astronautics

 a0 6= 0.
where a <qq is a positive definite learning rate, <+ is a design parameter and W
For scaling of the system consider (14) and

az (t) , a (t)/,
(37)
az0 , a0 /.

Then, by scaling the learning rate a = a0 /2 (36) can be written as

a (t) = a0 az (t)eT
W z (t)P B,
a (0) = W
W a0 , (38)

yielding invariance of the artificial weights update law with respect to the command coefficient , which is
also consistent with the update law for artificial basis functions, which can now be given as
aT (t)[B T B]1 B T {e z (t) Ar ez (t)},
az (t) = ka W az (0) = az0 , az0 6= 0. (39)
Downloaded by CARLETON UNIVERSITY LIBRARY on July 31, 2015 | http://arc.aiaa.org | DOI: 10.2514/6.2015-0608

Conclusively, (34,39) are scalable in the sense described in this paper. Note furthermore, that the adaptive
weight update law (35) is invariant to the command coefficient when scaling the learning rate = 0 /2

(t) = (x(t), c(t))eT (t)P B + (x(t), c(t)) T (t)W

 
W a (t) = (40)
a
 
= 0 z (x(t), c(t))eT T
z (t)P B + z (x(t), c(t))az (t)Wa (t) . (41)

Additionally, (33) can now be given as

 
= Ar z(t) + Br c0 (t) B W
z(t) T (t)z (x(t), c(t)) + W
aT (t)az (t) , z(0) = z0 , (42)

which renders scalability of the system equation. As a conclusion, scaling the learning rates of the adaptive
weight update laws (35,36) with 1/2 , adaptive controllers using artificial basis functions according to Ref.
13 can be scaled in the sense introduced in this paper.

III. Illustrative Examples

In this section, the scalability notion is illustrated using the controlled longitudinal motion of a Boeing 747
aircraft for numerical examples with reference model modification, e-modification, low-frequency learning,
artificial basis function based, and command governor based adaptive control architectures as introduced in
Sections C and D.

A. Simulation Model
Specifically, consider the uncertain dynamical system representing the controlled longitudinal motion of a
Boeing 747 airplane linearized at an altitude of 40 kft and a velocity of 774 ft/s given by

x(t)
= Ax(t) + Bu(t) + BW T (x(t), c(t)), x(0) = 0, (43)

where

0.003 0.039 0 0.322
0.065 0.319 7.74 0

A= , (44)
0.020 1.01 0.429 0
0 0 1 0
h iT
B= 0.010 0.180 1.16 0 , (45)

where x1 (t) represents the x-body-axis component of the velocity of the aircraft center of mass with respect
to the reference axes (in ft/s), x2 (t) represents the z-body-axis component of the velocity of the aircraft
center of mass with respect to the reference axes (in ft/s), x3 (t) represents the y-body-axis component of the
angular velocity of the aircraft (pitch rate) with respect to the reference axes (in rad/s), x4 (t) represents the

7 of 17

American Institute of Aeronautics and Astronautics

 pitch Euler angle of the aircraft body axes with respect to the reference axes (in rad), and u(t) represents
the elevator input (in rad). Here,
h i
WxT = 0.068 0.340 1.444 0 (46)

is an unknown ideal weight representing uncertainty due to modeling error in the pitch rate dynamics15 .
Additionally, a constant elevator bias of 0.03 is assumed, resulting in a total uncertainty of
h i
W T = 0.068 0.340 1.444 0 0.03 . (47)

 T
Furthermore, the regressor vector is given by (x(t), c(t)) = xT (t) 1 . For the simulations, we choose
the design matrix Q = I4 for the Lyapunov Equation (12) and the nominal control gains Kc = 8.9655 and
h i
Downloaded by CARLETON UNIVERSITY LIBRARY on July 31, 2015 | http://arc.aiaa.org | DOI: 10.2514/6.2015-0608

Kx = 0.0045 0.0759 9.6154 0.3546 . (48)

B. Simulations
For all examples the command profile for the pitch rate is given by c(t) = c (t), where

c (t) = 4 /s, 4s t < 8s, 12s t 16s, (49)

c (t) = 4 /s, 0s t < 4s, 8s t < 12s. (50)

The initial conditions are given by

 T
x0 = 0 0 3.5 /s 0 , (51)
 T
xr0 = 0 0 1.5 /s 0 (52)

for the system with = 1. For the second system, = 1.5 and the initial conditions are scaled as introduced
in Section II.B.

1. Reference Model Modification

As a first example, the reference model modification introduced in Section II.C.3 is used for the illustration
of the scalability notion with 0 = 0.5I5 and L = 4.5I4 .
Figures 1 and 2 show the response of the system with = 1 and = 1.5, respectively. The system
responses with different scaling command coefficients can be seen to be qualitatively identical. This is
illustrated further in Figure 3, where it is shown that scaling the first system with = 1.5 and plotting it
with the second system yields an identical response. Additionally, subtracting the scaled first system from
the second system shows a difference of numerical magnitude, which is displayed in Figure 4. Thus, the
qualitative observation from Figures 1 to 3 is also numerically verified. However, if the standard reference
model modification architecture is used without the scaling factor, the responses using the same learning
rate and different command do not display a scaled response, as expected. This is illustrated by Figure 5
with large scaled errors in both the system states and input.
Furthermore, Figure 6 displays that the adaptive weights are not invariant to the command coefficient
when the scaling factor is not used and hence, the scaled errors can be explained by the sensitivity of the
adaptive weights response to the overall level of excitation as a combination of the excitation of the regressor
vector (x(t), c(t)) and the learning rate . Note that just the first three components of the adaptive
weights are used in the figures of all examples in order to increase readability. Finally, Figure 7 shows that
the adaptive weight response is invariant to the command coefficient as discussed in Section II.B when the
scaling factor is used, which is consistent with the scaled performance as shown in the figures above.
This numerical example highlights the key advantage of introducing the proposed scaling factor. By
scaling the learning rates predictable performance with respect to a well tuned nominal system response is
achieved. However, if not used the system response is not predictable for nonidentical commands.

8 of 17

American Institute of Aeronautics and Astronautics

 2. e Modification
As an example of the robustness modifications presented in Section II.C.1, e-modification is used with 0 = I5
and e = 50. This particular set of gains is chosen on purpose to show scalability for oscillatory responses.
Figure 8 shows the response of the system with = 1 and = 1.5, respectively, when the scaling factor
is used. As mentioned, the plot shows poor tracking performance due to the chosen set of gains. However,
the system responses show to be of qualitatively similar shape. When plotting the scaled comparison of the
two systems, this fact is further emphasized as displayed in Figure 9. Building the scaled difference as shown
in Figure 10, it is obvious that the error between the scaled system stays within numerical magnitude, which
shows that the scaling notion works for the e-modification.
As mentioned before, for a system to be scalable, it is required that the adaptive weights are invariant to
the scaling factor. Figure 11 shows that the responses of the adaptive weights are identical and thus, invariant
to the command coefficient. As a conclusion, adaptive control architectures employing e-modification can
be scaled in the sense introduced in this paper.
Downloaded by CARLETON UNIVERSITY LIBRARY on July 31, 2015 | http://arc.aiaa.org | DOI: 10.2514/6.2015-0608

3. Low-Frequency Learning
Now, the low-frequency learning adaptive control framework introduced in Section II.C.2 is used for an
example utilizing 0 = 12I5 , f = 0.2I5 and = 12. Figure 12 shows the response of the system with = 1
and = 1.5, respectively, when the scaling factor is used.
Figures 13 and 14 show the response of the adaptive weights W and the filtered adaptive weights Wf,
respectively, when the scaling factor is not used. Although some more oscillations can be identified for
the larger command coefficient, the overall tendency of the adaptive weights shows qualitatively improved
performance with respect to the reference model modification displayed in Figure 6. Furthermore, the
difference introduced by the different command coefficient remains small in comparison to the adaptive
weight response of the reference model modification. Figure 15 shows this difference between the adaptive
weights, which is comparatively small in particular for the filtered adaptive weights Wf.
However, Figure 16 emphasizes that the scaled difference between the adaptive weight responses stays
within errors of numerical magnitude when the scaling factor is utilized. Hence the adaptive weight update
law is invariant to the scaling factor and consequently, low-frequency learning adaptive controllers can be
scaled in the sense introduced in this paper.

4. Adaptive Control with Artificial Basis Functions

In this section, the scalability of the adaptive control framework employing artificial basis functions as shown
in Section II.D is illustrated. The simulation is executed for 0 = 0.1I5 , a0 = 0.1, = 1, and ka = 100.
The initial conditions are chosen a0 = 0.1 and a0 = 0.15, respectively, and W a0 = 0.1.
Figure 17 shows the response of the system with = 1 and = 1.5, respectively, when using the original
adaptive framework with artificial basis functions without scaling the learning rates . Figure 18 illustrates
the approximate scaling behavior of the adaptive control framework using artificial basis functions when the
first system is scaled with the command coefficient and plotted with the second system. The approximate
scaling property is further emphasized in Figure 19, showing a small scaled difference.
However, Figure 20 displays that the adaptive weights and the artificial weights are not invariant to
the scaling factor. Furthermore, Figure 21 illustrates that the artificial basis functions show significant
scaling errors. As a conclusion, employing artificial basis functions for adaptive control yields approximate
scalability of the uncertain systems responses by adapting the artificial basis functions and the artificial
weights accordingly.
Now, the scaling factor is applied on both, artificial weights update law and adaptive weights update law.
Figure 22 displays the fact, that the artificial basis functions are now scaled. Furthermore, the artificial and
adaptive weights are invariant to the command coefficient as shown in Figure 23. As a consequence, the
overall systems response is scaled as illustrated in Figure 24, where the scaled errors stay within numerical
magnitude when utilizing the scalability notion.
Note that the adaptive framework using artificial basis functions13 can be scaled as shown in Section II.D.
Although the behavior of this class of controllers is already approximately scalable by design, introducing
the scaling factor further yields exact scalability in the sense introduced in this paper.

9 of 17

American Institute of Aeronautics and Astronautics

 5. Command Governor
As a final example, the command governor framework as presented in Section II.C.4 is considered. The
command governor gain = 15 is used and the learning rate is chosen 0 = 0.1I5 .
Note that the command governor is designed for shaping the transient performance10 and hence, the
uncertain systems response is approximately scalable for different command coefficients. Figure 25 shows
the response of the system with = 1 and = 1.5, respectively, when using the original command governor
framework without scaling the learning rates . The system responses appear to be scaled. Figure 26
illustrates the approximate scalability in more detail, showing a small scaled difference between the two
uncertain systems when using the original command governor framework. Note that these small errors
imply better scalability of the command governor with respect to the adaptive control framework using
artificial basis functions as shown in the section above. This could be due to the fact that more differential
equations impose their nonlinear effects on the systems performance in case of the artificial basis functions.
Note that although the command governor shows approximately scalable responses, the adaptive weights
are not invariant to the scaling factor as displayed in Figure 27. Finally, Figure 28 displays that the scaled
Downloaded by CARLETON UNIVERSITY LIBRARY on July 31, 2015 | http://arc.aiaa.org | DOI: 10.2514/6.2015-0608

differences stay within numerical magnitude when utilizing the scalability notion. Consequently, command
governor based adaptive controllers are scalable in the sense introduced in this paper. Note further that
introducing a metric to evaluate approximate scalability may be an important step towards validation and
verification of such systems.

IV. Conclusion
This paper presented the scalability notion for adaptive control design using a model of the longitudinal
motion of a Boeing 747 aircraft. Analysis employing a variety of simulations illustrated the positive effect
on predictability of the closed-loop system responses of various adaptive control architectures. Scaling the
learning rates was shown to render scaled system responses for systems with linear regressor vectors,
giving the opportunity to extend the performance of a well-tuned case to diverse nonidentical, but scalable
command profiles. Future work will include extensions of the scalability notion to nonlinear regressor vectors,
achieving aproximate scalability, and research on metrics evaluating scalability for existing adaptive control
architectures.

References
1 S. P. Schatz and T. Yucelen, Scalability Concept for Predictable Closed-Loop Response of Adaptive Controllers,

arXiv:1409.1695, 2014
2 K. S. Narendra and A. M. Annaswamy, Stable Adaptive Systems. Mineola, NY: Dover Publications, 2005.
3 K. J. Astrom and B.Wittenmark, Adaptive Control. Upper Saddle River, NJ: Prentice-Hall, 1994.
4 K. A. Wise, E. Lavretsky, and N. Hovakimyan, Adaptive control of flight: theory, applications, and open problems,

American Control Conference, 2006.

5 E. N. Johnson and A. J. Calise, Limited Authority Adaptive Flight Control for Reusable Launch Vehicles, Journal of

Guidance, Control, and Dynamics, vol. 26, 2003.

6 T. Yucelen and W. M. Haddad, Low-frequency learning and fast adaptation in model reference adaptive control, IEEE

Transactions on Automatic Control, 2013.

7 N. Hovakimyan and C. Cao, L1 adaptive control theory : Guaranteed robustness with fast adaptation. Philadelphia, PA:

Society for Industrial and Applied Mathematics, 2010.

8 E. Lavretsky, Reference dynamics modification in adaptive controllers for improved transient performance, AIAA

Guidance, Navigation, and Control Conference, 2011.

9 T. E. Gibson, A. M. Annaswamy, and E. Lavretsky, Adaptive systems with closed-loop reference models: Stability,

robustness, and transient performance, IEEE Transactions on Automatic Control (submitted).

10 T. Yucelen and E. N. Johnson, A new command governor architecture for transient response shaping, International

Journal of Adaptive Control and Signal Processing, 2013.

11 P. Ioannou and P. Kokotovic, Instability analysis and improvement of robustness of adaptive control, Automatica, vol.

20, pp. 583594, 1984.

12 K. S. Narendra and A. M. Annaswamy, A new adaptive law for robust adaptation without persistent excitation, IEEE

Transactions on Automatic Control, vol. 32, pp. 134145, 1987.

13 T. Yucelen and E. N. Johnson, Artificial Basis Functions in Adaptive Control for Transient Performance Improvement,

AIAA Guidance, Navigation, and Control Conference, 2013.

14 B. Gruenwald, T. Yucelen and A. Albattat, Performance Optimization of Adaptive Control Architectures, IEEE

Conference on Decision and Control, 2014

15 A. E. Bryson, Control of Aircraft and Spacecraft. Princeton, NJ: Princeton University Press, 1993.

10 of 17

American Institute of Aeronautics and Astronautics

 Figure 1. Response of reference system and the uncertain system using reference model modification when
= 1 (where the dotted lines indicate the system response).
Downloaded by CARLETON UNIVERSITY LIBRARY on July 31, 2015 | http://arc.aiaa.org | DOI: 10.2514/6.2015-0608

Figure 2. Response of reference system and the uncertain system using reference model modification when
= 1.5 when the scaling factor is utilized (where the dotted lines indicate the system response).

Figure 3. Scaled comparison of the system response using reference model modification with = 1.5 (solid
line) and = 1 (dotted line) when the scaling factor is utilized.

Figure 4. Scaled difference in system responses using reference model modification when the scaling factor is
utilized.

11 of 17

American Institute of Aeronautics and Astronautics

 Figure 5. Scaled difference in system responses using reference model modification when the scaling factor is
not utilized.
Downloaded by CARLETON UNIVERSITY LIBRARY on July 31, 2015 | http://arc.aiaa.org | DOI: 10.2514/6.2015-0608

Figure 6. Adaptive weights as a function of time using reference model modification when = 1 (left) and
= 1.5 (right) when the scaling factor is not utilized.

Figure 7. Adaptive weights as a function of time using reference model modification when = 1 (left) and
= 1.5 (right) when the scaling factor is utilized.

Figure 8. Response of reference system (solid line) and the uncertain system (dotted line) using e-modification
when = 1 (left) and = 1.5 (right), and the scaling factor is utilized.

12 of 17

American Institute of Aeronautics and Astronautics

 Figure 9. Scaled comparison of the system response using e-modification with = 1.5 (solid line) and = 1
(dotted line) when the scaling factor is utilized.
Downloaded by CARLETON UNIVERSITY LIBRARY on July 31, 2015 | http://arc.aiaa.org | DOI: 10.2514/6.2015-0608

Figure 10. Scaled difference in system responses using e-modification when the scaling factor is utilized.

Figure 11. Adaptive weights using e-modification as a function of time when = 1 (left) and = 1.5 (right),
and the scaling factor is utilized.

Figure 12. Response of reference system (solid line) and the uncertain system (dotted line) using low frequency
learning when = 1 (left) and = 1.5 (right), and the scaling factor is utilized.

13 of 17

American Institute of Aeronautics and Astronautics

 Figure 13. Response of adaptive weights using low frequency learning when = 1 (left) and = 1.5 (right)
and the scaling factor is not utilized.
Downloaded by CARLETON UNIVERSITY LIBRARY on July 31, 2015 | http://arc.aiaa.org | DOI: 10.2514/6.2015-0608

Figure 14. Response of filtered adaptive weights using low frequency learning when = 1 (left) and = 1.5
(right) and the scaling factor is not utilized.

Figure 15. Difference of filtered adaptive weights (left) and adaptive weights (right) using low frequency
learning when the scaling factor is not utilized.

Figure 16. Difference of filtered adaptive weights (left) and adaptive weights (right) using low frequency
learning when the scaling factor is utilized.

14 of 17

American Institute of Aeronautics and Astronautics

 Figure 17. Response of reference system (solid line) and the uncertain system (dotted line) using artificial
basis functions when = 1 (left) and = 1.5 (right), and the scaling factor is not utilized.
Downloaded by CARLETON UNIVERSITY LIBRARY on July 31, 2015 | http://arc.aiaa.org | DOI: 10.2514/6.2015-0608

Figure 18. Scaled comparison of the system response with = 1.5 (solid line) and = 1 (dotted line) using
artificial basis functions when the scaling factor is not utilized.

Figure 19. Scaled difference in system responses using artificial basis functions when the scaling factor is not
utilized.

Figure 20. Difference in adaptive weights (right) and artificial weights (left) as a function of time when the
scaling factor is not utilized.

15 of 17

American Institute of Aeronautics and Astronautics

 Figure 21. Difference (left) and scaling of artificial basis functions (right) as a function of time when the
scaling factor is not utilized (where the dotted line represents the scaled first system).
Downloaded by CARLETON UNIVERSITY LIBRARY on July 31, 2015 | http://arc.aiaa.org | DOI: 10.2514/6.2015-0608

Figure 22. Difference (left) and scaling of artificial basis functions (right) as a function of time when the
scaling factor is utilized (where the dotted line represents the scaled first system).

Figure 23. Difference in adaptive weights (right) and artificial weights (left) as a function of time when the
scaling factor is utilized.

Figure 24. Scaled difference in system responses using artificial basis functions when the scaling factor is
utilized.

16 of 17

American Institute of Aeronautics and Astronautics

 Figure 25. Response of reference system (solid line) and the uncertain system (dotted line) using original
command governor when = 1 (left) and = 1.5 (right).
Downloaded by CARLETON UNIVERSITY LIBRARY on July 31, 2015 | http://arc.aiaa.org | DOI: 10.2514/6.2015-0608

Figure 26. Scaled difference in system responses using original command governor.

Figure 27. Adaptive weights using original command governor as a function of time when = 1 (left) and
= 1.5 (right).

Figure 28. Scaled difference in system responses using command governor when the scaling factor is utilized.

17 of 17

American Institute of Aeronautics and Astronautics