Adaptive Control for Distributed Leader-Following

Consensus of Multi-Agent Systems

A Thesis
Presented to
the Faculty of the School of Engineering and Applied Science
University of Virginia

In Partial Fulfillment
of the Requirements for the Degree
Master of Science in Electrical Engineering

by

Ge Song

May 2015
APPROVAL SHEET
This thesis is submitted in partial fulfillment of the
requirements for the degree of
Master of Science in Electrical Engineering

Ge Song, Author

This thesis has been read and approved by the examining committee:

Professor Gang Tao, Thesis Advisor

Professor Zongli Lin

Professor Maite Brandt-Pearce

Accepted for the School of Engineering and Applied Science:

Dean, School of Engineering and Applied Science

May, 2015
Acknowledgments

First, I would like to gratefully thank my advisor, Professor Gang Tao. Without
his patient guidance and persistent encouragement, I would never completed this
thesis. His deep knowledge and insight on control theory helps me to solve problems
from one to another. It is a great privilege to have studied under him through the
past two years.
Second, I would also like to extend my gratitude to my research group members
and friends at the University of Virginia who share ideas with me.
Finally, I would like to sincerely thank my parents for their help and love. Without
their continuous and unconditional support during my time in University of Virginia,
I would never got the opportunity to be the person I am today.
 i

Abstract

The Distributed leader-following consensus problem for multi-agent systems has

drawn increasing attention recently. Consensus is a fundamental approach for dis-
tributed coordination. It means that a group of agents are made to reach an agree-
ment on some common states using certain local information. In the leader-following
consensus problem, there exists an active leader which specifies the movement of the
whole group. A majority of existing research is focused on the leader-following con-
sensus problem assuming that the parameters of follower agents are uncertain, while
few papers consider the leader dynamic uncertainty at the same time.
This thesis studies the distributed leader-following consensus problem of multi-
agent systems in which the leader and followers both have parametric uncertainties
and bounded external disturbances. Follower agents are controlled to follow an active
leader with a reference input signal, despite such uncertainties. An Adaptive control
method is adopted to solve this problem. This research starts from the basic case
that there are one leader and one follower in a multi-agent system. A new adaptive
scheme is proposed for dealing with parametric uncertainties. Furthermore, in order
to cancel the effect of disturbances, an adaptive disturbance compensator is developed.
Then, expanding the size of the multi-agent system under a directed graph, a new
distributed control protocol only using local information is adopted, generalizing the
previous control scheme. The proposed distributed control protocol has the capability
to guarantee that all agents can reach an agreement asymptotically with disturbances
acting on the follower agents. Comparing with the classical fixed gain control method,
the adaptive control method is capable of effectively handling system ans disturbance
uncertainties. Extensive numerical simulation results illustrate the effectiveness of
the proposed adaptive control scheme.
Contents

1 Introduction 1
1.1 Research Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Background 6
2.1 Control System Models . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Signal Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 System Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.4 Classical Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.5 Adaptive Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.6 Graph Theory and Communication Topology . . . . . . . . . . . . . . 16

3 Adaptive Leader-Following Control 18

3.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.1.1 Adaptive Control of Single Input Systems . . . . . . . . . . . 19
3.1.2 Adaptive Control of Multiple Inputs Systems . . . . . . . . . . 21
3.2 Adaptive Following Control Design for Single Input Systems . . . . . 23
3.2.1 Design Conditions . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2.2 Adaptive Control Laws . . . . . . . . . . . . . . . . . . . . . . 24

ii
 iii

3.2.3 Stability Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 26

3.2.4 Disturbance Rejection . . . . . . . . . . . . . . . . . . . . . . 28
3.3 Adaptive Following Control Design with Multiple Inputs . . . . . . . 34
3.3.1 Design Conditions . . . . . . . . . . . . . . . . . . . . . . . . 34
3.3.2 Adaptive Control Scheme . . . . . . . . . . . . . . . . . . . . 36
3.3.3 Stability Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.3.4 Disturbance Rejection . . . . . . . . . . . . . . . . . . . . . . 40
3.3.5 Design Based on LDU Parametrization . . . . . . . . . . . . . 45
3.4 Simulation Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.4.1 Simulation Study for Single Input Systems . . . . . . . . . . . 52
3.4.2 Simulation Study with Multiple Inputs . . . . . . . . . . . . . 60
3.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

4 Adaptive Leader-Following Consensus for Multiple Agents 73

4.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.2 Algebraic Graph Theory . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.3 Adaptive Control Design . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.3.1 Design Conditions . . . . . . . . . . . . . . . . . . . . . . . . 77
4.3.2 Adaptive Control Scheme . . . . . . . . . . . . . . . . . . . . 85
4.4 Stability Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.5 Disturbance Rejection . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.6 Simulation Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93

5 Conclusions and Future Work 100

5.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
5.2 Future Research Topics . . . . . . . . . . . . . . . . . . . . . . . . . . 101
List of Figures

2.1 Block diagram of a system with output feedback and an adjustable

preamplifier gain k. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 State space representation of a plant with state feedback. Reproduced
from [29]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3 An illustrative of a multi-agent system with five followers and one leader 17

3.1 Follower state (solid) vs. leader state (dotted) with fixed gain control. 54
3.2 Follower state (solid) vs. leader state (dotted) with adaptive gain control. 54
3.3 Tracking errors with adaptive control vs. fixed gain control. . . . . . 55
3.4 Parameter errors ki (t) − ki∗ with adaptive control (i = 1, 2, 3). . . . . 55
3.5 Follower state (solid) vs. leader state (dotted) without disturbance
rejection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.6 Follower state (solid) vs. leader state (dotted) with adaptive distur-
bance rejection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.7 Tracking errors with adaptive disturbance rejection vs. without dis-
turbance rejection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.8 Follower state (solid) vs. leader state (dotted) without disturbance
rejection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

iv
 v

3.9 Follower state (solid) vs. leader state (dotted) with adaptive distur-
bance rejection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.10 Tracking errors with adaptive disturbance rejection vs. without dis-
turbance rejection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.11 Follower state (solid) vs. leader state (dotted) with fixed gain control. 62
3.12 Follower state (solid) vs. leader state (dotted) with adaptive gain control. 62
3.13 Tracking errors with adaptive control vs. fixed gain control. . . . . . 63
3.14 Parameter errors Ki (t) − Ki∗ with adaptive control (i = 1, 2, 3). . . . . 63
3.15 Follower state (solid) vs. leader state (dotted) without disturbance
rejection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.16 Follower state (solid) vs. leader state (dotted) with adaptive distur-
bance rejection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.17 Tracking errors with adaptive disturbance rejection vs. without dis-
turbance rejection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.18 Follower state (solid) vs. leader state (dotted) without disturbance
rejection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.19 Follower state (solid) vs. leader state (dotted) with adaptive distur-
bance rejection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.20 Tracking errors with adaptive disturbance rejection vs. without dis-
turbance rejection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.21 Follower state (solid) vs. leader state (dotted) with adaptive gain con-
trol (LDU decomposition). . . . . . . . . . . . . . . . . . . . . . . . . 71
3.22 Follower state (solid) vs. leader state (dotted) with adaptive gain con-
trol (LDU decomposition). . . . . . . . . . . . . . . . . . . . . . . . . 71
3.23 Tracking errors with adaptive control vs. fixed gain control (LDU
decomposition). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
 vi

4.1 Case I: Interaction graph of two followers and one leader aaaaa aaaaaaaa 79
4.2 Case II: Interaction graph of two followers and one leader . . . . . . . 80
4.3 Case III: Interaction graph of two followers and one leader . . . . . . 80
4.4 Case IV: Interaction graph of two followers and one leader . . . . . . 80
4.5 A basic structure exists in directed graphs . . . . . . . . . . . . . . . 80
4.6 Interaction graph after adding one follower on follower v1 . . . . . . . 81
4.7 Interaction graph after adding one follower on follower v1 . . . . . . . 82
4.8 Interaction graph after adding one follower for follower vk . . . . . . . 83
4.9 Interaction graph of three follower agents and one leader . . . . . . . 94
4.10 State trajectories of three follower states xi and leader state x0 vs.
time(sec) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.11 Tracking errors between three follower states xi and leader state x0 vs.
time(sec) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.12 Interaction graph of five follower agents and one leader . . . . . . . . 96
4.13 State trajectories of five follower states xi and leader state x0 vs.
time(sec) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.14 Tracking errors between five follower states xi and leader state x0 vs.
time(sec) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.15 State trajectories of three follower states xi and leader state x0 with
disturbance rejection vs. time(sec) . . . . . . . . . . . . . . . . . . . 99
4.16 Tracking error between follower states xi and leader state x0 vs. time(sec) 99
Chapter 1

Introduction

In past two decades, an unprecedented growth in sensing, communications and

computation has been witnessed. This growth change the way we collecting and
processing the information. The sensor network revolution makes it possible to us for
exploring and interacting with the environment. Hence the growing importance of the
multi-agent system cooperative systems has also been acknowledge by the research
community. Many important scientific results have appeared addressing different
aspects of multi-agent cooperative systems. Due to the multi-disciplinary nature of
multi-agent cooperative systems, research focusing on this field can address many
parallel problems for instance in the area of distributed decision making, in the area
of connectivity maintenance and in the area of vehicle formation, etc. [32].
In the meanwhile, adaptive control method is becoming popular in many fields.
Since the control community has heavily acknowledged the importance of the adaptive
control, lots of useful results on adaptive control appearing to make this modern
control method more powerful. Hence, in this thesis, we combine these two popular
topics, multi-agent cooperative systems and the adaptive control method, together.
We will explore the leader-following consensus problems of multi-agent system by

1
 2

using adaptive control in the coming several chapters.

1.1 Research Motivation

With the development of sensor networks, control of multi-agent systems has been
emerging and has drawn lots of attention. Since the main objective in distributed
control of multi-agent systems is to enable a group of agents to perform a special
task, distributed control is also referred as the cooperative control. There are several
typical problems existing in the cooperative control including the consensus problem
[9, 37, 45], the formation problem [4, 8, 34] and the flocking problem [36, 39, 41].
Among those three typical cooperative control problems, the consensus problem
is the most important one because the consensus algorithm is basic and fundamen-
tal [26]. In particular, consensus means to reach an agreement on a certain quantity
of interest, namely their position or velocities [37]. As the most fundamental and
important control problem in cooperative control, Jadbabaie et al. [13] considered
such the leader-following consensus problem and proved that if all the agents were
jointly connected with their leader, their states would coverage to that of the leader as
time goes on. The controllabilty of the leader-following dynamic network was studied
in [14, 22, 46]. The leader-following consensus problem can be classified into two dif-
ferent kinds. The first one is the consensus regulation problem, which is also referred
as leaderless problem (behavior-based) [20, 27, 31, 35]. The other one is the consen-
sus tracking problem, also known as leader-following consensus(synchronization). For
leader-following consensus problems, there should exist an active leader whose states
keep changing in the multi-agent system [11, 21, 40, 50]. In another word, leader-
following consensus means that all follower agents eventually reach an agreement on
the state or output of a preassigned leader, which specifies a desired objective for all
 3

other agents to follow and is usually independent of its followers [24].

Formation control requires agents keeping a desired special distance, in another
word, formation control requires agents keeping a desired formation configuration
which actually is the consensus of relative position and flocking control requires agents
to move together with the same velocity and avoid inter-collisions at the same time
which obviously is the consensus on velocities [25]. The practical application of co-
operative control is broadly used in many areas such as formation control of mobile
vehicles and scheduling of automated highway systems.

1.2 Literature Review

In current leader-following consensus relative literature, distributed protocols are

designed for multi-agent systems with single-integrator or double-integrator linear
dynamics followers with an active leader. The active leader model is also governed by
a single-integrator or double-integrator linear dynamics. A plenty of literature focuses
on single-integrator and/or double-integrator systems like [9, 10, 33]. For example,
[33], it considered a consensus algorithm for double-integrator dynamics with several
different cases. In [51], the consensus problem was addressed for a double integrator
multi-agent system to track a single-integrator leader with a desired constant velocity,
and meanwhile the non-uniform time-varying communication delays were taken into
account. Recently, the multi-agent system with general linear dynamics has also
been considered. Recent design and analysis tools cover from the output regulation
approach in [11] and [40] which based on the output regulation theory [3, 5, 48]. [28]
investigated the consensus leader-following problem for a group of identical followers
with an autonomous active leader whose parameter matrix is as same as the follower.
In [28] only local information can be used by the distributed controller. Being different
 4

with [28], [24] solved the consensus problem for a group of followers with different
dynamics under an external disturbances with an active leader which model can be
also different with the followers. Plant parameters of the followers in [24] are in
uncertainty. The distributed control protocol in [24] designed by model reference
adaptive control method and also uses the local information of an individual agent.
However, although the author said that some part of the leader’s parameter can be
unknown, this statement is not convinced. Motivations of the aforementioned research
are clear but in practical application sometimes conditions of those researches are
unsatisfied.
Motivated by the above observations, in this thesis we study the adaptive leader-
following consensus problem of a group of linear dynamics agents guided by an active
real leader with external disturbances under directed graphs. Different from [28], sys-
tem parameter matrices in this thesis can be completely different with each other, and
also the system parameter matrices of the leader are supposed to be different from all
the followers. Comparing to [24], dynamic parameters can be completely unknown
including dynamic parameters of all followers and the leader. This assumption is
meaningful because in the real engineering environment, sometimes it is difficult to get
accurate dynamic parameters. In the meanwhile, like many papers [12, 17–19, 23, 44],
this thesis also considers the external disturbances acting on the real application.
Hence, we develop a new protocol structure based on the principle of adaptive con-
trol method in order to solve the consensus leader-following problem with an external
disturbances such that all follower agents can track the active leader with a desired
reference signal. Then, the performance of a close-loop system is analyzed to verify
the designed control protocol with the corresponding adaptive laws. Finally, sev-
eral simulation results will be displayed to verify the effectiveness of our proposed
theoretical protocol.
 5

1.3 Thesis Outline

The remainder of this thesis is organized as follows. In Chapter 2 we introduce

some basic background information to help readers follow the thesis easier including
background about control systems and background about multi-agent systems. In
Chapter 3, we develop a new adaptive control scheme for leader-following consensus
problem with only one leader-one follower with and without disturbances and present
simulation results to support the new adaptive scheme. In Chapter 4, we present
a new adaptive control scheme for multi-agent systems with directed graph whose
structure is modified from the control scheme developed in Chapter 3. Several cases
are presented including the disturbance-free case and the disturbance-acting case. The
corresponding numerical simulation results illustrate the effectiveness of the adaptive
control protocol. Finally, in Chapter 5 we discuss the results in this thesis and the
potential future work.
Chapter 2

Background

Before starting to discuss the adaptive control of multi-agent system, some basic
background about control system and multi-agent system needs to be presented. In
this chapter, topics about control system including control system modeling, system
stability, classical control and adaptive control are presented first. Topics following
the background of control system are some basic background about topology.

2.1 Control System Models

Usually in order to solve an engineering problem, the first step is to develop an

appropriate mathematical model of the system either from physical laws or from
experimental data. Dynamic systems are described by differential equations. While
most dynamics systems are nonlinear in nature, study of their linearized models and
linear systems has played a crucial role in understanding dynamic system behaviors.
In the first coming paragraph, a topic to be discussed is the non-linear system. In
the second coming paragraph the linear system is discussed.

6
 7

Nonlinear Systems A dynamic system can be described by a set of differential

equations

Fi (y (p) (t), . . . , y (1) (t), y(t), u(p) (t), . . . , u(1) (t), u(t), t) = 0, t ≥ t0 , (2.1)

di y(t) di u(t)
i = 1, 2, . . . , l. y (i) (t) and u(i) (t) denote the ith time derivatives dti
and dti
of
y(t) and u(t), with a common notation ẏ(t) = y (1) (t), u̇(t) = u(1) (t) and ÿ(t) =
y (2) (t), ü(t) = u(2) (t). A special form of Fi depends on a special system under consid-
eration.
A n-th order dynamic system can be described by a group of n interactive first-
order differential equations

ẋ = f0 (x, u, t), y = h0 (x, u, t), t ≥ t0 , (2.2)

for some functions f0 ∈ Rn and h0 ∈ Rq with q ≥ 1, n ≥ 1, where x = [x1 , . . . , xn ]T ∈

Rn is the system state vector with the state variable xi , i = 1, 2, . . . , n, physical or
artificial, to completely define the system behavior, u(t) ∈ RM with M ≥ 1 is the
system input, and y(t) is the system output [42].
The system behavior depends on the control u(t). When a feedback control law
u(t) = β(x, t) is use, system (2.2) becomes

ẋ = f (x, u, t), y = h(x, u, t), t ≥ t0 , (2.3)

for some functions f ∈ Rn and h ∈ Rq with q ≥ 1, n ≥ 1, whose solution is denoted

as x(t) = x(t; t0 , x0 ), where x(t0 ) = x0 is the initial state vector. For a vector function
x(t), a measure of its “magnitude” is its norm kxk [42].
However, sometimes in order to make system analysis easier, we need to degenerate
 8

the non-linear systems to the linear system on (x0 , u0 ) by Taylor expansion,1 which
is

δ ẋ = Aδx + Bδu

δy = Aδx + Bδu (2.4)

with
   
∂f1 ∂f1 ∂f1 ∂f1
 ∂x1 ... ∂xn   ∂u1 ... ∂un 
∂f  .

.. 
 ∂f  .

.. 

A= |(x ,u ) =  .. .  , B= |(x ,u ) =  .. .  ,
∂x 0 0   ∂u 0 0  
   
∂fn ∂fn ∂fn ∂fn
∂x1
... ∂xn ∂u1
... ∂un
|(x0 ,u0 ) |(x0 ,u0 )
   
∂h1 ∂h1 ∂h1 ∂h1
 ∂x1 ... ∂xn   ∂u1 ... ∂un 
∂h  .

.. 
 ∂h  .

.. 

C= |(x ,u ) =  .. .  ,D = |(x ,u ) =  .. .  .
∂x 0 0   ∂u 0 0  
   
∂hn ∂hn ∂hn ∂hn
∂x1
... ∂xn ∂u1
... ∂un
|(x0 ,u0 ) |(x0 ,u0 )

.
Since linearization always be made in a small range, in oder to denote conveniently
usually people express linearized system (2.4) as

ẋ = Ax + Bu

y = Cx + Du (2.5)

(2.5) is called the “linearized system” and the method used in this linearization pro-
cess is called ‘Lyapunov linearization method”. Lyapunov linearization method is
concerned with the local stability of a nonlinear system. It is a formalization of the
intuition that a nonlinear system method should behave similarly to its linearized
1
Often, (x0 , u0 ) is an equilibrium point which means f (x0 , u0 ) = 0
 9

approximation for small range motions. Because all physical systems are inherently
nonlinear, Lyapunov’s linearizion method serves as the fundamental justification of
using linear control techniques in practice. More details about Lyapunov linearization
method can be found in literature [38].

Linear Systems As we discussed in the previous paragraph, many nonlinear sys-

tems are hard to be analyzed or to be controlled. One of the methods to control
nonlinear systems is linearization. Thus there is no doubt that researches on linear
system are really important.
A linear system time-varying dynamic system is described as

ẋ = A(t)x + B(t)u

y = C(t)x + D(t)u, t ≥ t0 , (2.6)

with x(t0 ) = x0 ,where x ∈ Rn , u ∈ RM , y ∈ Rq are the state, input and output,

respectively. A(t) ∈ Rn×n , B(t) ∈ Rn×M , C(t) ∈ Rq×n and D(t) ∈ Rq×M are continu-
ous functions of t. Systems like (2.6) are linear time varying system. If parameters
A, B, C ans D in (2.6) are constant which does not depends on t, the system is called
the “linear time invariant (LTI) system”. In the time-invariant case, system (2.6)
becomes

ẋ = Ax + Bu

y = Cx + Du, t ≥ t0 , (2.7)

where A ∈ Rn×n , B ∈ Rn×M , C ∈ Rq×n , and D ∈ Rq×M are all constant system
 10

matrices. In this case, the state solution can be explicitly expressed as

Z t
A(t−t0 )
x(t) = e x0 + eA(t−τ ) Bu(τ )dτ (2.8)
t0

Gain matrix decomposition Leader-following systems to be discussed in this the-

sis are LTI systems. Usually, in control input signal, there exists one or more gains
making the control input signal work. Introducing an important system characteri-
zations for linear system which is useful to the later control design development.
Assume that all leading principal minors of the gain matrix Kp are nonzero 2 .
The LDU decomposition of Kp exists and can be employed for adaptive control of
the plant (2.7) [42].
Let ∆i , i = 1, 2, . . . , M be the leading principle minors of Kp , The following
proposition gives the key result of the LDU decomposition.

Proposition 2.1. (LDU decomposition) A matrix Kp ∈ RM ×M with all its leading

principle minors being nonzero has a unique decomposition:

Kp = LD∗ U (2.9)

for some M × M unit(i.e., with all diagonal elements being 1) lower triangular matrix
L and unit upper triangular matrix U , and

 
∗ ∆2 ∆M
D = diag {d∗1 , d∗2 , . . . , d∗M } = diag ∆1 , ,..., . (2.10)
∆1 ∆M −1

This is the well-known LDU decomposition of a nonsingular matrix with nonzero

 
a11 ... a1k
 . .. 
2
The kth leading principal minor of a matrix A = aij ∈ RM ×M  ..
is det  . ,k =


ak1 ... akk

1, 2, . . . , M.
 11

leading principal minors.

2.2 Signal Measures

In order to have a unique measurement for vector signals, a new concept called
“norm” is defined. Consider a vector signal x(t) = [x1 (t), . . . , xn (t)]T ∈ Rn . x(t) =
[x1 (t), . . . , xn (t)]T is a vector at any t, and is a vector function as t changes. Vector
norms can measure vectors, and while function norms can measure vector functions.

Definition 2.1. A real-valued function k·k on linear space S is a norm if

(i) kxk ≥ 0 for all x ∈ S and kxk = 0 only if x = 0;
(ii) kβxk = |β| kxk for all x ∈ S and any scalar β; and
(iii) kx + yk ≤ kxk + kyk for all x, y ∈ S.

Since in control systems there exist many signal vectors that need to be measured,
it is imperative to introduce signal norms before we develop a adaptive control scheme.
The L1 , L2 , and L∞ norms are defined below as

L1 = {x(t) ∈ Rn : kx(·)k1 < ∞}, (2.11)

L2 = {x(t) ∈ Rn : kx(·)k2 < ∞}, (2.12)

L∞ = {x(t) ∈ Rn : kx(·)k∞ < ∞}, (2.13)

where the vector signal norms are

 12

Z ∞
kx(·)k1 = (|x1 (t)| + · · · + |xn (t)|)dt, (2.14)
0
sZ
∞
kx(·)k2 = x21 (t) + · · · + xn1 (t), (2.15)
0

kx(·)k∞ = sup max |xi (t)|. (2.16)

t≥0 1≤i≤n

Remark 2.1 By definition of L∞ , we can conclude that x(t) ∈ L∞ iff x(t) is bounded,
i.e., x(t) ∈ L∞ and x(t) is bounded are equivalent. 

2.3 System Stability

The concept of stability is crucial to control system design. An unstable control

system is useless and dangerous. The methods available to examine the poles depend
on the representation of the system model. If the classical approach is taken then the
poles of the transfer function can be examined. If the modern approach is used then
the eigenvalues, which are the poles, of the system matrix A can be analyzed. Either
approach can quickly give information on whether or not the system is inherently
stable, marginally stable, or unstable.
For adaptive control systems stability must be defined another way since knowl-
edge of the system parameters are unavailable and possibly changing. The work of
Alexander Mikhailovich Lyapunov, who presented definitions and theorems for study-
ing the stability of solutions to a broad class of differential equations, has been used
extensively to address this problem [16]. The work of Lyapunov relies on defining an
energy function, formally known as a Lyapunov function candidate, that can be used
to determine the stability of a system without having to solve for the solutions to the
system explicitly. Originally, this Lyapunov function was purely the total mechanical
 13

or electrical energy and therefore by nature positive definite.

Lyapunov indirected method can be found in many textbooks about nonlinear
system like [42], [38].

Stability definitions Since all systems to be discussed in this thesis are LTI sys-
tems, we will introduce the definition of stability for LTI [2]

Definition 2.2. The response of ẋ(t) = Ax(t) is marginally stable or stable in the
sense of Lyapunov if every finite initial state x0 excites a bounded response. It is
asymptotically stable if every finite state excites a bounded response which, in addition,
approaches 0 as t → ∞.

Usually, we do not use the definition to check the stability of a LTI system.
Theorem 2.1 can help us to check the stability of a LTI system more quickly.

Theorem 2.1. The equation ẋ(t) = Ax(t) is marginally stable if and only if all eigen-
values of A have zero or negative real parts and those with zero real parts are simple
roots of the minimal polynomial of A. The equation ẋ(t) = Ax(t) is asymptotically
stable if and only if all eigenvalues of A have negative real parts.

Also in the Lyapunov sense, we can check the stability of matrix A by Lyapunov
theorem

Theorem 2.2. The equation ẋ = Ax, A ∈ Rn×n , x ∈ Rn is asymptotically stable

if and only if for every positive definite Q = QT ∈ Rn×n , the Lyapunov equation
AT P + P A = −Q has a unique and positive definite solution P = P T ∈ Rn×n .

Theorem 2.1 and Theorem 2.2 are theorems which we usually used to check the
stability of close-looped system by classical control. But for adaptive control, The-
orem 2.1 and Theorem 2.2 would not work because in adaptive control there exist
 14

uncertainty on the dynamics model, we have no access to get a set of accurate system
parameters. So we introduce a new method called Lyapunov direct method to help
us check the system stability when applying adaptive control.

Theorem 2.3. (Lyapunov direct method) If in some Ball B(h) there exists a positive
definite function V (x, t) with V̇ ≤ 0, then the equilibrium state xe = 0 of the the
system ẋ = Ax is stable. If, in addition, V (x, t) is decrescent, then it is uniformly
stable.

Based on definitions and theorems presented before, Barbalart lemma is intro-

duced. This lemma makes the precess to analyze system stability more easier.

Lemma 2.1. (Barbalat Lemma)If a scalar function f˙(t) ∈ L∞ , f (t) ∈ L2 , then

limt→ f (t) = 0. [42]

2.4 Classical Control

The first thing we need to know about the classical control is the feedback is
pervasive. Feedback is a very crucial method to stabilize the unstable system stable
[1]. Usually, the output y(t) is fed back and compared with the input u(t). The block
diagram of the most classical feedback control system is shown in Fig 2.1. System
transfer function can be computed by block diagrams like Fig 2.1. Remind that in Fig
2.1, k is a constant through the whole control process. Whether control parameter k
can update or not is one of a big differences with adaptive control and the classical
control.
With the state space representation it is convenient to feedback the state variables
x(t) to the control signal u(t) instead of an input signal y(t). With this configuration
each state variable can be adjusted by a gain vector K to give the desired closed loop
 15

u(t) y(t)
k G(s)
−

Figure 2.1: Block diagram of a system with output feedback and an adjustable pream-
plifier gain k.

poles. A typical control system represented with the state space representation uti-
lizing state feedback is displayed in Figure 2.2 [29] where double error lines represent
vector signals.

2.5 Adaptive Control

Unlike classical control systems, adaptive control system has capability to payload
system uncertainties including unknown system structures, system parameters and
other environmental uncertainties. An adaptive controller structure is designed based
on a known parameter case, however, the value of the adaptive controller parameter is
updated by parameter estimator instantly. Combining an adaptive controller with an
parameter estimators, adaptive control can achieve a desired performance by dealing
with variety kind of uncertainties.
There are two approaches to adaptive control design. The first one is direct adap-

r(t) + u(t) + Ẋ R X y(t)

B C
+ +

−K

Figure 2.2: State space representation of a plant with state feedback. Reproduced
from [29].
 16

tive control. For direct adaptive control, the parameter estimator estimates controller
parameters online directly according to the plant information of input/output. Plant
parameters are parametrized implicitly in terms of a set of parameters of a nominal
controller. The second approach is indirected adaptive control and is characterized
by estimating the parameters of the plant first. Then based on some design equation
to calculate control parameters to achieve the desired performance. Direct adaptive
control will be employed in this thesis.

2.6 Graph Theory and Communication Topology

Information exchange between agents can be represented as a graph. We give

some basic terminology and definitions from graph theory [6, 7] which will be used in
Chapter 4.

Definition 2.3. By a graph G we mean a finite set V(G) = (vi , . . . , vn ), whose ele-
ments are called nodes or vertexes, together with set E(G ⊂ V × V), whose elements
are called edges. An edge is therefore an ordered pair of distinct vertexes.

Definition 2.4. A graph is called direct graph if and only if for all (vi , vj ) ∈ E(G),
the edge (vj , vj ) ∈ E(G), then the graph is said to be undirected. Otherwise, it is called
a directed graph.

Definition 2.5. An edge (vi , vj ) is said to incoming with respect to vj and outgoing
with respect to vi ) and can be represented as an arrow with vertex vi as its tail and
vertex as its head.

Definition 2.6. A path of length r in a directed graph is a sequence (v0 , . . . , vr ) of

r + 1 distinct vertexes such that for every i ∈ {0, . . . , r − 1}, (vi , vi+1 ) is an edge.
 17

2 3 4

0 1 5
Figure 2.3: An illustrative of a multi-agent system with five followers and one leader

aaaaaa
aaaaaa
aaaaaa
aaaaaa
aaaaaa
aaaaaa
aaaaaa
aaaaaa
aaaaaa
aaaaaa
aaaaaa
Chapter 3

Adaptive Leader-Following Control

The final objective for this thesis is to design a distributed adaptive control scheme
of leader-following consensus problem for multi-agent system. However, before going
that far, it is important for us to know how adaptive control works for systems in-
cluding one follower and one leader. Fully understand this basic problem is important
for us to develop a more complex controller for multi-agent systems in Chapter 4.

3.1 Problem Statement

For state tracking control problems, there are two different common cases we are
interested in. One is the control problem with a single input, the other one is the
control problem with multiple inputs. Adaptive control of single input is easier and
more fundamental than multiple input cases and adaptive control of multiple input
systems are more significant and practical. In this chapter, both of these two control
problems will be discussed.

18
 19

3.1.1 Adaptive Control of Single Input Systems

Consider a linear time-invariant follower plant in state-space form

ẋ(t) = Ax(t) + bu(t) + bd(t), x(t) ∈ Rn , u(t) ∈ R, t ≥ 0 (3.1)

with x(0) = x0 , where A ∈ Rn×n is an unknown constant parameter matrix, and

b ∈ Rn is an unknown constant parameter vector and d(t) is an external unknown
bounded disturbance. The state x(t) ∈ Rn is measurable. The leader dynamic system
is give by

ẋm (t) = Am xm (t) + bm um (t) + bm dm (t), xm (t) ∈ Rn , um (t) ∈ R, t ≥ 0 (3.2)

where Am is an unknown constant parameter matrix and bm is an unknown constant

parameter vector. xm (t) and um (t) are available for measurement and bounded. dm (t)
is an bounded disturbance acting on the leader dynamic system.
Regarding by the disturbances d(t) and dm (t), there are three different cases in
total.

Case I: Single input systems without disturbances In this case, d(t) = 0 and
dm (t) = 0. There exist no disturbance in the multi-agent system. The dynamic of
the follower and the leader can be re-presented as

ẋ(t) = Ax(t) + bu(t), x(t) ∈ Rn , u(t) ∈ R, t ≥ 0 (3.3)

ẋm (t) = Am xm (t) + bm um (t), xm (t) ∈ Rn , um (t) ∈ R, t ≥ 0 (3.4)

 20

Case II: Single input systems with a disturbance acting on the followers
In this case, d(t) 6= 0 and dm (t) = 0. The dynamic plant of the follower is

ẋ(t) = Ax(t) + bu(t) + bd(t), x(t) ∈ Rn , u(t) ∈ R, t ≥ 0 (3.5)

d(t) ∈ R is an unknown external disturbance acting on the follower system. The

disturbance should be bounded and satisfies

q
X
d(t) = d0 + dβ fβ (t) (3.6)
β=1

where d0 and dβ are unknown constant and fβ (t) is a bunch of known continuous
bounded functions for β = 1, 2, . . . , q for some q ≥ 0. The dynamic of the given leader
remains as (3.4).

Case III: Single input systems with disturbances acting on both the leader
and the followers In this case, d(t) 6= 0 and dm (t) 6= 0. When d(t) and dm (t) are
not equal zero, the follower plant is (3.5). The leader dynamic is

ẋm (t) = Am xm (t) + bm um (t), xm (t) ∈ Rn , um (t) ∈ R, t ≥ 0 (3.7)

where
qm
X
dm (t) = dm0 + dmβ fmβ (t) (3.8)
β=1

with dm0 and dmβ are unknown constant and fmβ (t) are some known continuous
bounded functions for β = 1, 2, . . . , p for some p ≥ 0.
The control objective is to design a bounded state feedback control signal u(t) to
make the follower system state x(t) bounded and tracking the leader system xm (t)
asymptotically, i.e., limt→∞ (x(t) − xm (t)) = 0.
 21

3.1.2 Adaptive Control of Multiple Inputs Systems

In section 3.1.1, the control input u(t) and the leader reference signal um (t) are
scalars, and parameters b and bm are vectors. However, in practical engineering
applications, there are many multiple inputs systems with u ∈ Rm , m 6= 1.
Consider a follower plant expressed as (3.9), and the leader is given as (3.10)

ẋ(t) = Ax(t) + Bu(t) + Bd(t), x(t) ∈ Rn , u(t) ∈ Rp (3.9)

ẋm (t) = Am xm (t) + Bm um (t) + Bm dm (t), xm (t) ∈ Rn , um (t) ∈ Rm (3.10)

where A ∈ Rn×n is an unknown parameter matrix and Am ∈ Rn×n is an asymp-

totically stable parameter matrix. However, B ∈ Rn×p and Bm ∈ Rn×m become to
parameter matrices rather than parameter vectors. Although the dimension of B and
Bm are changed, all the parameters matrices here including A, Am , B and Bm remain
unknown in this section. xm (t) and um (t) are available for measurement and bounded
and x(t) is measured.
Being similar to single input system, classify multiple inputs systems into three
cases regarding by the existence of the disturbances.

Case I: Multiple inputs systems without disturbances d(t) = 0 and dm (t) =

0. In this situation, the one leader-one follower system degenerate into

ẋ(t) = Ax(t) + Bu(t), x(t) ∈ Rn , u(t) ∈ Rp , t ≥ 0 (3.11)

ẋm (t) = Am xm (t) + Bm um (t), xm (t) ∈ Rn , um (t) ∈ Rm , t ≥ 0 (3.12)

Case II: Multiple inputs systems with a disturbance acting on the follower
systems d(t) 6= 0 and dm (t) = 0. The dynamic plant of the follower is
 22

ẋ(t) = Ax(t) + Bu(t) + Bd(t), x(t) ∈ Rn , u(t) ∈ Rp , t ≥ 0 (3.13)

d(t) ∈ Rp is an unknown external disturbance acting on the follower system. The

disturbance should be bounded and satisfies

q
X
d(t) = d0 + dβ fβ (t) (3.14)
β=1

where d0 ∈ Rp and dβ ∈ Rp are unknown constant vectors and fβ (t) are a bunch of
known continuous bounded functions for β = 1, 2, . . . , q for some q ≥ 0. The given
leader remains as

ẋm (t) = Am xm (t) + Bm um (t), xm (t) ∈ Rn , um (t) ∈ Rm , t ≥ 0 (3.15)

Case III: Single input systems with disturbance acting on both leader and
follower d(t) 6= 0 and dm (t) 6= 0. the follower plant is (3.13). The dynamic of the
given leader is

ẋm (t) = Am xm (t) + Bm um (t) + Bm dm (t), xm (t) ∈ Rn , um (t) ∈ Rm , t ≥ 0 (3.16)

where
q
X
dm (t) = dm0 + dmβ fmβ (t) (3.17)
β=1

with dm0 ∈ Rm and dmβ ∈ Rm are unknown constant and fmβ (t) are some known
continuous bounded functions for β = 1, 2, . . . , q for some p ≥ 0.
The control objective is to design a bounded control signal u(t) to make the
follower system state x(t) bounded and track the leader system state xm (t) asymp-
totically.
 23

3.2 Adaptive Following Control Design for Single

Input Systems

In this section, we will solve the single input system leader-following state tracking
problem. The single input system without disturbance will be solved first which in-
cludes the basic principle of the leader-following state tracking problem. Disturbance
rejection is developed in Section 3.2.4.

3.2.1 Design Conditions

As stated in Section 3.1.1, the control objective is to design a bounded state

feedback control signal u(t) to make the follower system state x(t) bounded and
tracking the leader system xm (t) asymptotically.
To meet the control objective, we assume

(A3.1) all the eigenvalues of Am are in the open left-half complex plane;

(A3.2) um (t) is bounded and piecewise continuous;

(A3.3) there exist two constant vectors k1∗ ∈ Rn and k3∗ ∈ Rn and a non-zero constant
scalars k2∗ ∈ R, k4∗ ∈ R such that the following equations are satisfied:

Ae = A + bk1∗T , bm = bk2∗

Am = A + bk3∗T , be = bk4∗ , (3.18)

where Ae ∈ Rn×n is a stable and known matrix and be ∈ Rn is a known vector;

(A3.4) sign[k4∗ ], the sign of the parameter k4∗ , is known;

 24

Assumptions (A3.1) and (A3.2) are for a stable, well-defined reference system with
a bounded output ym (t). Assumption (A3.3) is the so-called matching condition such
that if the parameters of Am , bm , A and b are known and (3.18) is satisfied then the
control law
u∗ (t) = k1∗T (x(t) − xm (t)) + k2∗ um (t) + k3∗T xm (t) (3.19)

achieves the control objectives: the closed-loop system becomes

ẋ(t) = Ax(t) + b(k1∗T (x(t) − xm (t)) + k2∗ um (t) + k3∗T xm (t))

= (A + bk1∗T )(x(t) − xm (t)) + (A + bk3∗T )xm (t) + bk2∗ um (t)

= Ae (x(t) − xm (t)) + Am xm (t) + bm um (t) (3.20)

whose state vector x(t) belongs to L∞ , and the tracking error e(t) = x(t) − xm (t)
satisfies:
ė(t) = Ae e(t), e(0) = x(0) − xm (0) (3.21)

which indicates that limt→∞ e(t) = 0 exponentially. So that with the nominal con-
troller (3.19) , the desired control objective can be achieved.

3.2.2 Adaptive Control Laws

In our problem, the parameters of A, b, Am and bm are unknown, so (3.19) can not
be used for control. In this case an adaptive controller which has the same structure
of (3.19) is to be used, whose structure is given as

u(t) = k1T (t)(x(t) − xm (t)) + k2 (t)um (t) + k3T (t)xm (t) (3.22)
 25

where k1 (t), k2 (t) and k3 (t) are the estimates of k1∗ , k2∗ and k3∗ respectively. The de-
sign task now is to choose adaptive laws update these estimates so that the control
objective is still achievable even all the four parameters are unknown.
To be specific, the adaptive laws to update the control parameters are proposed
as

k̃˙ 1 (t) = k̇1 (t) = −sign[k4∗ ]Γe(t)eT (t)P be (3.23)

k̃˙ 2 (t) = k̇2 (t) = −sign[k4∗ ]γum (t)eT (t)P be (3.24)

k̃˙ 3 (t) = k̇3 (t) = −sign[k4∗ ]Ψxm (t)eT (t)P be (3.25)

where P ∈ Rn×n is a positive definite matrix satisfying ATe P + P Ae = −Q, for any
chosen Q ∈ Rn×n being constant and Q = QT > 0. Γ ∈ Rn×n , Ψ ∈ Rn×n are constant
matrices, and Γ = ΓT > 0, Ψ = ΨT > 0, γ > 0 is a constant scalar. k1 (0), k2 (0) and
k3 (0) are arbitrary.
In summary, according to the previous development, we now present the following
result.

Theorem 3.1. The adaptive controller (3.22), with the adaptive laws (3.23), (3.24)
and (3.25), applied to the plant (3.3) guarantees that all closed-loop signals are bounded
and the tracking error e(t) = x(t) − xm (t) goes to zero as t goes to infinity, i.e.,
limt→∞ e(t) = 0.
 26

3.2.3 Stability Analysis

To prove Theorem 3.1, we see

u(t) = u(t) − u∗ (t) + u∗ (t)

= (k1 (t) − k1∗ )T (x(t) − xm (t)) + (k2 (t) − k2∗ )um (t)

+ (k3 (t) − k3∗ )T xm (t) + u∗ (t), (3.26)

and then we derive the adaptive control based on the tracking error equation

ė(t) = Ae e(t) + b(k1 (t) − k1∗ )T e(t) + b(k2 (t) − k2∗ )um (t) + b(k3 (t) − k3∗ )T xm (t)
 
1 T 1 1 T
= Ae e(t) + be k̃ (t)e(t) + ∗ k̃2 (t)um (t) + ∗ k̃3 (t)xm (t) (3.27)
k4∗ 1 k4 k4

where
k̃1 (t) = k1 (t) − k1∗ , k̃2 (t) = k2 (t) − k2∗ , k̃3 (t) = k3 (t) − k3∗ (3.28)

are parameter errors.

Since the adaptive laws for k1 (t), k2 (t) and k3 (t) are chosen to be dynamics from
some adaptive laws, the state vector of the closed-loop error system is

ec (t) = (eT (t), k̃1T (t), k̃2 (t), k̃3T (t))T ∈ R3n+1 . (3.29)

We choose a positive definite function as a Lyapunov function candidate

1 T −1 1 1
V (ec ) = eT P e + ∗ 1
k̃ Γ k̃1 + ∗ k̃22 γ −1 + ∗ k̃3T Ψ−1 k̃3 (3.30)
|k4 | |k4 | |k4 |

as a measurement of the system errors. As stated in the adaptive laws (3.23)-(3.25),

 27

P ∈ Rn×n is positive definite and satisfies:

ATe P + P Ae = −Q < 0. (3.31)

Parameter matrices Γ and Ψ are symmetric positive definite, and parameter γ is a

positive scalar as aforementioned in the previous Subsection 3.2.2.
Compute the time derivative of V (ec )

d
V̇ = V (ec )
dt
 T  T    T
∂V (ec ) ∂V (ec ) ˙ ∂V (ec ) ˙ ∂V (ec )
= ė(t) + k̃1 (t) + k̃2 (t) + k̃˙ 3 (t)
∂e ∂ k̃1 ∂ k̃2 ∂ k̃3
2 2 2
= 2eT (t)P ė(t) + ∗ k̃1T (t)Γ−1 k̃˙ 1 (t) + + ∗ k̃2 (t)γ −1 k̃˙ 2 (t) + ∗ k̃3T (t)Ψ−1 k̃˙ 3 (t)
|k4 | |k4 | |k4 |
(3.32)

Substituting (3.27) and (3.31) in (3.32), we have

2 T 2
V̇ = −eT (t)Qe(t) + eT (t)P be k̃ (t)e(t) + eT (t)P be ∗ k̃2 (t)um (t)
∗ 1
k4 k4
2 2 2
+ eT (t)P be ∗ k̃3T (t)xm (t) + ∗ k̃1T Γ−1 k̃˙ 1 (t) + ∗ k̃2 (t)γ −1 k̃˙ 2 (t)
k4 |k4 | |k4 |
2
+ ∗ k̃3T (t)Ψ−1 k̃˙ 3 (t) (3.33)
|k4 |

By the adaptive laws, (3.23)-(3.25), (3.33) becomes

V̇ = −eT (t)Qe(t) ≤ −qm ke(t)k22 ≤ 0 (3.34)

where qm > 0 is the minimum eigenvalue of Q. From here on, the desired properties
of the proposed adaptive laws are obvious:

(i) V > 0 and V̇ ≤ 0 implies that the equilibrium state ec = 0 of the closed-
 28

loop system consisting of (3.27), (3.23), (3.24) and (3.25) is uniformly stable
and its solution ec (t) is uniformly bounded, which gives the boundedness of
x(t), k1 (t), k2 (t) and k3 (t), and in turn of the boundedness of ė(t) because of
(3.27);

(ii) (3.34) implies e(t) ∈ L2 ;

(iii) with e(t) ∈ L2 ∩ L∞ and ė(t) ∈ L∞ , applying Barbalat lemma, we conclude that
limt→∞ e(t) = 0. 5

From the results demonstrated above, all the properties mentioned in Theorem 3.1
are proved.

3.2.4 Disturbance Rejection

In Section 3.1.1 Case II has a disturbance acting on the follower, and Case III has
disturbances acting on the leader and the followers respectively. To reject the effect
of unknown disturbances so that the desired system performance can be achieved,
certain matching conditions should be satisfied and additional compensation term
should be introduced to the controller (3.22).

Design for the disturbance acting on the follower system(Case II) When
Assumptions (A3.1)-(A3.4) are satisfied, the ideal control law (3.19) is modified as

u(t) = k1∗ (x(t) − xm (t)) + k2∗ um (t) + k3∗ xm (t) + k5∗ (t), (3.35)

Pq
where k5∗ = −d(t) = −d0 − β=1 dβ fβ (t), which leads to the desired closed-loop
system
ẋ(t) = Ae (x(t) − xm (t)) + Am xm (t) + bm um (t) (3.36)
 29

an in turn the desired tracking error system ė(t) = Ae e(t) making limt→∞ e(t) = 0,
as Ae is stable.
When the parameters A, b, Am , bm and the value of d(t) are unknown, the adaptive
version of the controller (3.35) is used:

u(t) = k1T (t)(x(t) − xm (t)) + k2 (t)um (t) + k3T (t)xm (t) + k5 (t), (3.37)

where ki (t) for i = 1, 2, 3 are the estimates of ki∗ , respectively, and

q
X
k5 (t) = k50 (t) + k5β (t)fβ (t) (3.38)
β=1

∗
with k5β (t) being the estimate of k5β = −dβ , β = 0, 1, 2, · · · , q.
To develop adaptive laws for ki (t) for i = 1, 2, 3 and k5β (t), β = 0, 1, 2, . . . , q, we
first derive an system error equation in terms of the tracking error e(t) = x(t) − xm (t)
and the parameter errors

k̃i (t) = ki (t) − ki∗ , i = 1, 2, 3,

∗
k̃5β (t) = k5β (t) − k5β , β = 0, 1, 2, . . . , q. (3.39)

Let k̃5 (t) = k5 (t) − k5∗ . Using (3.37),(3.38) and (3.39), we obtain

ẋ(t) = Ax(t) + b(k1T (t)(x(t) − xm (t)) + k2 (t)um (t) + k3T (t)xm (t) + k5 (t))

= Ae (x(t) − xm (t)) + Am xm (t) + bm um (t)

1 T
+ be (k̃ (t)(x(t) − xm (t)) + k̃2 (t)um (t) + k̃3T xm (t) + k̃5 (t)) (3.40)
k4∗ 1
 30

Substituting (3.7) into (3.40), we have the tracking error equation

1 T
ė(t) = Ae e(t) + be (k̃ (t)(x(t) − xm (t)) + k̃2 (t)um (t) + k̃3T xm (t) + k̃5 (t)). (3.41)
k4∗ 1

Based on this error equation, we choose the adaptive laws as

k̇1 (t) = −sign[k4∗ ]Γ1 e(t)eT (t)P be , (3.42)

k̇2 (t) = −sign[k4∗ ]γ2 um (t)eT (t)P be , (3.43)

k̇3 (t) = −sign[k4∗ ]Ψ3 xm (t)eT (t)P be , (3.44)

k̇50 (t) = −sign[k4∗ ]γ50 eT (t)P be , (3.45)

k̇5β (t) = −sign[k4∗ ]γ5β fβ (t)eT (t)P be , β = 1, 2, · · · , q, (3.46)

where Γ1 = ΓT1 > 0, γ2 > 0, Ψ3 = ΨT3 > 0 and γ5β > 0, β = 0, 1, 2, · · · , q are
adaptation gains, and P = P T > 0 satisfies (3.31).
To analysis the close-loop system stability, choose a positive definite function as
Lyapunov candidate function which is
q
1 1 2 −1 1 1 X 2 −1
V = e P e + ∗ k̃1T Γ−1
T
1 k̃1 + k̃ γ + ∗ k̃3T Ψ−1
∗ 2 2 3 k̃3 + k̃ γ (3.47)
|k4 | |k4 | |k4 | |k4∗ | β=0 5β 5β

Taking time derivative and apply the adaptive laws (3.42) - (3.46) into its derivative.
The time derivative of (3.47)

2 T 2
V̇ = −eT (t)Qe(t) + eT (t)P be k̃ (t)e(t) + eT (t)P be ∗ k̃2 (t)um (t)
∗ 1
k4 k4
q
T 2 T T 2 X 2
+ e (t)P be ∗ k̃3 (t)xm (t) + e (t)P be ∗ k̃5β (t) + eT (t)P be ∗ k̃5β (t)fβ (t)
k4 k4 β=1
k4
2 T −1 ˙ 2 2
+ ∗ 1
k̃ Γ k̃1 (t) + ∗ k̃2 γ −1 k̃˙ 2 (t) + ∗ k̃3T Ψ−1 k̃˙ 3 (t)
|k4 | |k4 | |k4 |
q
2 −1 ˙
X 2
+ ∗ k̃50 γ50 k̃50 (t) + k̃ γ −1 k̃˙ (t)
∗ 50 5β 5β
(3.48)
|k4 | β=1
|k4 |
 31

By the adaptive laws (3.42)-(3.46), (3.48) becomes to

V̇ = −eT (t)Qe(t), Q = QT > 0 (3.49)

from which we can obtain the desired system properties, i.e., we can also have the
result of Theorem 3.1. 5
Remark 3.1 The disturbance d(t) in (3.3) is matched to the control input u(t), that
is, both act on the plant dynamics through the same vector b. If they are not matched,
for example, for the plant

ẋ(t) = Ax(t) + bu(t) + bd (d0 + d1 f (t)), (3.50)

where b and bd ∈ Rn are linear independent, the above adaptive design may not be
able to ensure asymptotic state tracking. 

Design for disturbance acting on both leader and follower systems (Case
III) For the purpose of designing an adaptive control scheme to address the dis-
turbance acting on both leader and follower systems, we assume all the Assumptions
(A3.1)-(A3.4) aforementioned are satisfied. In order to reject the disturbance so that
the desired system performance can be achieved, the ideal controller with disturbance
compensator is chosen as:

u(t) = k1∗T (x(t) − xm (t)) + k2∗ um (t) + k3∗T xm (t) + k5∗ (t) (3.51)

which has the same form with the ideal controller chosen in the last paragraph.
However, when d(t) and dm (t) present in (3.6) and (3.7) respectively at the same
 32

time, we re-define

k5∗ = k2∗ dm (t) − d(t)

qm q
X X
= (k2∗ dm0 − d0 ) + k2∗ dmβ fmβ (t) − dβ fβ (t)
β=1 β=1
qm q
X X
∗ ∗ ∗
= k50 + k5mβ fmβ (t) + k5β fβ (t) (3.52)
β=1 β=1

with

∗
k50 = k2∗ dm0 − d0 , k5mβ
∗
= k2∗ dmβ , k5β
∗
= −dβ . (3.53)

The ideal controller (3.51) with the re-defined k5∗ leads (3.5) to a desired system
closed-loop system

ẋ(t) = Ae (x(t) − xm (t)) + Am xm (t) + bm um (t) + bm dm (t) (3.54)

In result, we obtain limt→∞ e(t) = 0, since ė(t) = ẋ(t) − ẋm (t) = Ae e(t) and Ae is
stable.
When the parameters A, b, Am , bm and the value of disturbance d(t) are unknown,
we use the adaptive version of controller (3.50) as

u(t) = k1T (t)(x(t) − xm (t)) + k2 (t)um (t) + k3T (t)xm (t) + k5 (t) (3.55)

where
qm q
X X
k5 (t) = k50 (t) + k5mβ (t)fmβ (t) + k5β (t)fβ (t) (3.56)
β=1 β=1

is the estimate of k5∗ .

To develop adaptive laws for ki (t) for i = 1, 2, 3, k5β (t) for β = 0, 1, 2, . . . , q and
 33

k5mβ (t) for β = 1, 2, . . . , p, derive the tracking error e(t) = x(t) − xm (t) first and the
parameter errors

k̃i (t) = ki (t) − ki∗ , i = 1, 2, 3,

∗
k̃5β (t) = k5β (t) − k5β , β = 0, 1, 2, . . . , q,
∗
k̃5mβ (t) = k5mβ (t) − k5mβ , β = 1, 2, . . . , qm . (3.57)

Let k̃5 (t) = k5 (t) − k5∗ . Substituting (3.50), (3.56) and (3.57) into (3.54), we obtain

1 T
ẋ(t) = Ae (x(t) − xm (t)) + be (k̃ (t)(x(t) − xm (t)) + k̃2 (t)um (t) + k̃3T xm (t) + k̃5 (t))
k4∗ 1
+ Am xm (t) + bm um (t) + bm dm (t) (3.58)

With e(t) = x(t) − xm (t), we have the tracking error equation

1 T
ė(t) = Ae e(t) + be (k̃ (t)(x(t) − xm (t)) + k̃2 (t)um (t) + k̃3T xm (t) + k̃5 (t)). (3.59)
k4∗ 1

Based on (3.59), we choose the adaptive laws as

k̇1 (t) = −sign[k4∗ ]Γ1 e(t)eT (t)P be , (3.60)

k̇2 (t) = −sign[k4∗ ]γ2 um (t)eT (t)P be , (3.61)

k̇3 (t) = −sign[k4∗ ]Ψ3 xm (t)eT (t)P be , (3.62)

k̇50 (t) = −sign[k4∗ ]γ50 eT (t)P be , (3.63)

k̇5β (t) = −sign[k4∗ ]γ5j fβ (t)eT (t)P be , β = 1, 2, · · · , qm , (3.64)

k̇5mβ (t) = −sign[k4∗ ]δ5mβ fmβ (t)eT (t)P be , β = 1, 2, · · · , p (3.65)

Analysis the closed-loop system stability by Lyapunov method. Choose a positive

 34

definite function as
q
1 T −1 1 1 1 X 2 −1
V = eT P e + k̃ Γ k̃1 + ∗ k̃22 γ2−1 + ∗ k̃3T Ψ−1
∗ 1 1 3 k̃3 + k̃ γ
|k4 | |k4 | |k4 | |k4∗ | β=0 5β 5β
qm
1 X 2 −1
+ ∗ k̃5mβ δ5mβ (3.66)
|k4 | β=1

has the negative semidefinite time derivative after substituting (3.60) -(3.65)

V̇ = −eT (t)Qe(t), Q = QT > 0 (3.67)

from which, we can obtain all the desired system properties, i.e, the control objective
is achieved. 5

3.3 Adaptive Following Control Design with Mul-

tiple Inputs

In this section, we focus on the leader-following consensus problem with multiple

inputs. Case I in Section 3.1.2 will be solved first. Case II and Case III in Section 3.1.2
with disturbances are developed in Section 3.3.4. In Section 3.3.5, an adaptive control
scheme is designed for multiple inputs systems based on the LDU decomposition.

3.3.1 Design Conditions

Recall that in Section 3.1.2, the control objective for the multiple inputs systems
is stated as to design a control input u(t) to make the follower system state x(t)
bounded and to track the leader state xm (t) asymptotically. In order to achieve this
control objective, several design conditions are presented as follows.

(A3.5) all the eigenvalues of Am are in the open left-half complex plane;
 35

(A3.6) um (t) is bounded and piecewise continuous;

(A3.7) there exist four parameter matrices K1∗ ∈ Rn×p , K2∗ ∈ Rp×m , K3∗ ∈ Rn×p and
K4∗ ∈ Rp×p such that the following equations are satisfied:

A + BK1∗T = Ae , BK2∗ = Bm , A + BK3∗T = Am , BK4∗ = Be , (3.68)

where Ae ∈ Rn×n is a stable and known matrix and Be ∈ Rn×p is a known

matrix;

(A3.8) there is a known matrix S ∈ Rp×p such that K4∗ S is symmetric and positive
definite: Ms = K4∗ S = (K4∗ S)T = S T K4∗T > 0.

If the parameter of Am , Bm , A and B are known, (3.68) is satisfied, then the ideal
control law
u∗ (t) = K1∗T (x(t) − xm (t)) + K2∗ um (t) + K3∗T xm (t) (3.69)

results in the closed-loop system

ẋ(t) = Ae (x(t) − xm (t)) + Am xm (t) + Bm um (t) (3.70)

whose state vector x(t) belongs to L∞ , i.e., x(t) is bounded. The tracking error
e(t) = x(t) − xm (t) satisfies:

ė(t) = Ae e(t), e(0) = x(0) − xm (0), (3.71)

so that limt→∞ e(t) = 0 exponentially since Ae is a stable matrix.

 36

3.3.2 Adaptive Control Scheme

When the parameters of A, b, Am and bm are unknown, update (3.69) to the adap-
tive version,

u(t) = K1T (t)(x(t) − xm (t)) + K2 (t)um (t) + K3T (t)xm (t) (3.72)

where K1 (t), K2 (t) and K3 (t) are the estimates of K1∗ , K2∗ and K3∗ respectively.
Choose the adaptive laws (3.73) - (3.75) to update K1 (t), K2 (t) and K3 (t):

K̃˙ 1T = K̇1T (t) = −S T BeT P e(t)eT (t) (3.73)

K̃˙ 2 = K̇2 (t) = −S T BeT P e(t)uTm (t) (3.74)

K̃˙ 3T = K̇3T (t) = −S T BeT P e(t)xTm (t) (3.75)

with S satisfying Assumption (A3.8) and P = P T > 0 satisfying ATe P +P Ae = −Q <

0 for any chosen symmetric positive definite matrix Q. K1 (0), K2 (0) and K3 (0) are
chosen arbitrarily.

Theorem 3.2. The adaptive controller (3.72) with the adaptive laws (3.73), (3.74)
and (3.75), applied to the system (3.11) guarantees that all closed-loop signals are
bounded and the tracking error e(t) = x(t) − xm (t) goes to zero as t goes to ∞.

3.3.3 Stability Analysis

In order to prove the Theorem 3.2, in another word, to prove that the closed-loop
system (3.11) with state feedback adaptive controller (3.72) is stable, we first define
the parameter error as

K̃1 (t) = K1 (t) − K1∗ , K̃2 (t) = K2 (t) − K2∗ , K̃3 (t) = K3 (t) − K3∗ . (3.76)
 37

Use (3.11) and (3.72) to obtain

ẋ(t) = Ax(t) + Bu(t)

= Ae (x(t) − xm (t)) + Am xm (t) + Bm um (t)

+ B(K̃1T (t)(x(t) − xm (t)) + K̃2 (t)um (t) + K̃3T (t)xm (t))

= Ae (x(t) − xm (t)) + Am xm (t) + Bm um (t)

 
∗−1 T ∗−1 ∗−1 T
+ Be K4 K̃1 (t)e(t) + K4 K̃2 (t)um (t) + K4 K̃3 (t)xm (t) . (3.77)

Substituting (3.12) in ė(t) = x(t) − xm (t), we have the tracking error equation

 
ė(t) = Ae e(t) + Be K4∗−1 K̃1T (t)e(t) + K4∗−1 K̃2 (t)um (t) + K4∗−1 K̃3T (t)xm (t) . (3.78)

Now K1 (t), K2 (t) and K3 (t) are matrices so the state vector of the closed-loop
system is

ec (t) = (eT (t), k̃11

T T
(t), . . . , k̃1n T
(t), k̃21 T
(t), . . . , k̃2m T
(t), k̃31 T
(t), . . . , k̃3n (t))T ∈ R(2n+m)p+n
(3.79)
where k̃1i (t) ∈ Rp is the ith column of K̃1T (t), i = 1, 2, . . . , n. k̃2j (t) ∈ Rp is the jth
column of K̃2 (t), j = 1, 2, . . . , m, and k̃3q (t) ∈ Rp is the qth column of K̃3T (t), q =
1, 2, . . . , n, that is,

K̃1T (t) = (k̃11 (t), . . . , k̃1n (t)) ∈ Rp×n ,

K̃2 (t) = (k̃21 (t), . . . , k̃2m (t)) ∈ Rp×m ,

K̃3T (t) = (k̃31 (t), . . . , k̃3n (t)) ∈ Rp×n . (3.80)

 38

We choose the positive definite function

n
X n
X n
X
T
V (ec ) = e P e + T
k̃1i Ms−1 k̃1i + T
k̃2j Ms−1 k̃2j + T
k̃3q Ms−1 k̃3q (3.81)
i=1 j=1 q=1

as a measurement of these errors, where P ∈ Rn×n is constant, P = P T > 0 and

satisfies (3.31) for some Q ∈ Rn×n being constant and Q = QT > 0, and Ms = MsT > 0
satisfies the Assumption A3.8. With tr[M ] denoting the trace of a square matrix M ,
we express V (ec ) as

V (ec ) = eT P e + tr[K̃1 Ms−1 K̃1T ] + tr[K̃2T Ms−1 K̃2 ] + tr[K̃3 Ms−1 K̃3T ] (3.82)

The time-derivative of V (ec ) is

 T n  T m
!T
∂V (ec ) ∂V (ec ) ˙ ∂V (ec )
k̃˙ 2j (t)
X X
V̇ = ė(t) + k̃1i (t) +
∂e i=1
∂ k̃ 1i j=1
∂ k̃2j
n
!T
∂V (ec )
k̃˙ 3q (t)
X
+
q=1 ∂ k̃3q
n n
(t)Ms−1 k̃˙ 1i (t) (t)Ms−1 k̃˙ 2j (t)+
X X
T T T
= 2e (t)P ė(t) + 2 k̃1i +2 k̃2j
i=1 j=1
n
(t)Ms−1 k̃˙ 3q (t)
X
T
2 k̃3q (3.83)
q=1

= 2eT (t)P ė(t) + tr[K̃1 (t)Ms−1 K̃˙ 1T (t)] + tr[K̃2T (t)Ms−1 K̃˙ 2 (t)] + tr[K̃3 (t)Ms−1 K̃˙ 3T (t)]

Substituting (3.78) and (3.31) in (3.83), we have

V̇ = −eT (t)Qe(t) + 2eT (t)P Be K4∗−1 K̃1T (t)e(t) + 2eT (t)P Be K4∗−1 K̃2 (t)um (t)

+ 2eT (t)P Be K4∗−1 K̃3T (t)xm (t) + 2 tr[K̃1 (t)Ms−1 K̃˙ 1T (t)] + 2 tr[K̃2T (t)Ms−1 K̃˙ 2 (t)]

+ 2 tr[K̃3 (t)Ms−1 K̃˙ 3T (t)]. (3.84)

 39

Using the definition Ms = K4∗ S = MsT > 0 and the properties that tr[M1 M2 ] =
tr[M2 M1 ], tr[M3 ] = tr[M3T ] for any matrices M1 , M2 and M3 of appropriate dimen-
sions, we obtain

eT (t)P Be K4∗−1 K̃1T (t)e(t)

= tr[eT (t)P Be K4∗−1 K̃1T (t)e(t)] = tr[eT (t)P Be SMs−1 K̃1T (t)e(t)]

= tr[e(t)eT (t)P Be SMs−1 K̃1T (t)] = tr[K̃1 (t)Ms−1 S T BeT P e(t)eT (t)] (3.85)

eT (t)P Be K4∗−1 K̃2 (t)um (t)

= tr[eT (t)P Be K4∗−1 K̃2 (t)um (t)] = tr[eT (t)P Be SMs−1 K̃2 (t)um (t)]

= tr[um (t)eT (t)P Be SMs−1 K̃2 (t)] = tr[K̃2T (t)Ms−1 S T BeT P e(t)uTm (t)] (3.86)

eT (t)P Be K4∗−1 K̃3T (t)xm (t)

= tr[eT (t)P Be K4∗−1 K̃3T (t)xm (t)] = tr[eT (t)P Be SMs−1 K̃3T (t)xm (t)]

= tr[xm (t)eT (t)P Be SMs−1 K̃3T (t)] = tr[K̃3 (t)Ms−1 S T BeT P e(t)xTm (t)] (3.87)

Apply adaptive laws (3.73), (3.74) and (3.75) into (3.84), combing with the facts
(3.86) and (3.87), we have,

V̇ = −eT (t)Qe(t) ≤ −qm ke(t)k22 ≤ 0 (3.88)

Hence the equilibrium state ec = 0 of the closed-loop system consisting of (3.73),

(3.74), (3.75) and (3.78) is uniformly stable and its solution ec (t) is uniformly bounded.
That is, y(t), K1 (t), K2 (t), K3 (t) and ė(t) all are bounded. Furthermore (3.88) im-
plies e(t) ∈ L2 and so limt→0 e(t) = 0. Theorem 3.2 is proved. 5
 40

3.3.4 Disturbance Rejection

To reject the effect of unknown disturbances so that the desired system perfor-
mance can be achieved, certain matching conditions should be satisfied and additional
compensation is introduced to the controller (3.69).
Case II and Case III in Section 3.1.2 are developed in the later two paragraphs
respectively.

Design for disturbance acting on follower systems (Case II) When Assump-
tions (A3.5)-(A3.8) are satisfied, the ideal control law (3.69) is modified as

u(t) = K1∗ (x(t) − xm (t)) + K2∗ um (t) + K3∗ xm (t) + k5∗ (t), (3.89)

where k5∗ = −d(t) ∈ Rp , which leads to the desired closed-loop system

ẋ(t) = (A + BK1∗T )(x(t) − xm (t)) + (A + BK3∗T )xm (t) + BK2∗ um (t) + Bk5∗ + Bd(t)

= Ae (x(t) − xm (t)) + Am xm (t) + Bm um (t) (3.90)

an in turn the desired tracking error system ė(t) = Ae e(t) making limt→∞ e(t) = 0,
as Ae is stable.
When the parameters of A, B, Am , Bm and the value of d(t) are unknown, update
the controller (3.89) to an adaptive version which is

u(t) = K1 (t)t (x(t) − xm (t)) + K2 (t)um (t) + K3T (t)xm (t) + k5 (t), (3.91)
 41

where Ki (t), i = 1, 2, 3 are the estimates of Ki∗ , respectively, and

q
X
k5 (t) = k50 (t) + k5β (t)fβ (t) (3.92)
β=1

∗
with k5β (t) being the estimate of k5β = −dβ for β = 0, 1, . . . , q.
To develop adaptive laws for Ki (t) for i = 1, 2, 3, and k5β (t) for β = 0, 1, 2, · · · , q,
firstly we derive an error equation in terms of the tracking error e(t) = x(t) − xm (t)
and the parameter errors

K̃i (t) = Ki (t) − Ki∗ , i = 1, 2, 3,

∗
k̃5β (t) = k5β (t) − k5β , β = 0, 1, 2, · · · , q. (3.93)

Substituting (3.89) into (3.13) with the definition of parameter errors, then we
obtain

ė(t) = Ae e(t) + Be K4∗−1 (K̃1T (t)(x(t) − xm (t)) + K̃2 (t)um (t) + K̃3T xm (t) + k̃5 (t)) (3.94)

where k̃5 (t) = k5 (t) − k5∗ .

Based on this error equation, we choose the adaptive laws as

K̃˙ 1T = K̇1T (t) = −S T BeT P e(t)eT (t) (3.95)

K̃˙ 2 = K̇2 (t) = −S T BeT P e(t)uTm (t) (3.96)

K̃˙ 3T = K̇3T (t) = −S T BeT P e(t)xTm (t) (3.97)

K̃˙ 50 = K̇50 (t) = −S T BeT P e(t) (3.98)

K̃˙ 5β = K̇5β (t) = −S T BeT P e(t)fβT (t) (3.99)

where S satisfies Assumption (A3.8) and P = P T > 0 satisfies (3.31).

 42

Applying adaptive control (3.91) with adaptive laws (3.95) - (3.99), (3.13) should
have all the desired the properties. In order to show that the controller (3.91) with
adaptive laws (3.95) - (3.99) works and can drive e(t) = x(t) − xm (t) bounded and
stable which are the desired properties of trajectories of (3.13). Choose a positive
definite function

V = eT P e + tr[K̃1 Ms−1 K̃1T ] + tr[K̃2T Ms−1 K̃2 ] + tr[K̃3 Ms−1 K̃3T ]

q
X
+ T
tr[k̃50 Ms−1 k̃50 ] + T
tr[k̃5j Ms−1 k̃5j ]. (3.100)
β=1

Then the time derivative of (3.97) is

V̇ = −eT (t)Qe(t) + 2eT (t)P Be K4∗−1 K̃1T (t)e(t) + 2eT (t)P Be K4∗−1 K̃2 (t)um (t)

+ 2eT (t)P Be K4∗−1 K̃3T (t)xm (t) + 2eT (t)P Be K4∗−1 k̃50
T
(t)

+ 2eT (t)P Be K4∗−1 k̃5j f5j + 2 tr[K̃1 (t)Ms−1 K̃˙ 1T (t)]

+ 2 tr[K̃2T (t)Ms−1 K̃˙ 2 (t)] + 2 tr[K̃3 (t)Ms−1 K̃˙ 3T (t)]

T
+ 2 tr[k̃50 Ms−1 k̃˙ 50 ] + 2 tr[k̃5j
T
Ms−1 k̃˙ 5j ]. (3.101)

Applying adaptive laws (3.95)-(3.99), the time derivative of (3.101) as

V̇ = −eT (t)Qe(t) ≤ 0 (3.102)

which is negative semidefinite. From (3.102) we can obtain the desired system prop-
erties. 5

Disturbance acting on both leader and follower systems(Case III) Consider

the leader plant (3.13) and follower plant (3.16). In order to reject the disturbance
so that the desired system performance can be achieved, we choose the ideal control
 43

structure as:

u∗ (t) = K1∗T (x(t) − xm (t)) + K2∗ um (t) + K3∗T xm (t) + k5∗ (t). (3.103)

Being different with Case II, when d(t) and dm (t) is present in (3.13) and (3.16)
respectively, we re-define

k5∗ = K2∗ dm (t) − d(t)

qm q
X X
= (K2∗ dm0 − d0 ) + K2∗ dmβ fmβ (t) − dβ fβ (t)
β=1 β=1
qm q
X X
∗ ∗ ∗
= k50 + k5mβ fmβ (t) + k5β fβ (t) (3.104)
β=1 β=1

with

∗
k50 = K2∗ dm0 − d0 , k5mβ
∗
= K2∗ dmβ , k5β
∗
= −dβ (3.105)

For the ideal controller, the re-defined k5∗ leads to a desired closed-loop system

ẋ(t) = (A + BK1∗T )(x(t) − xm (t)) + (A + BK3∗T )xm (t) + BK2∗ um (t)

qm q
X X
+ BK2∗ dm0 − Bd0 + B K2∗ dmβ fmβ (t) −B dβ fβ (t) + Bd(t)
β=1 β=1

= Ae (x(t) − xm (t)) + Am xm (t) + Bm um (t) + Bm dm (t) (3.106)

In result, we obtain limt→∞ e(t) = 0, since ė(t) = ẋ − ẋm = Ae e(t) and Ae is stable.
When the parameters A, B, Am , Bm and the value of disturbance d(t) are unknown,
we use the adaptive version of controller (3.103) which is

u(t) = K1T (x(t) − xm (t)) + K2 um (t) + K3T xm (t) + k5 (t) (3.107)

 44

where
qm q
X X
k5 (t) = k50 (t) + k5mβ (t)f5mβ (t) + k5β (t)fβ (t) (3.108)
β=1 β=1

is the estimate of k5∗ .

In adaptive control, it is unaccessible for us to obtain the ideal parameters, so
define the parameter errors first.

K̃i (t) = Ki (t) − Ki∗ , i = 1, 2, 3.

∗
k̃5β (t) = k5β (t) − k5β , β = 0, 1, . . . , q.
∗
k̃5mβ (t) = k5mβ (t) − k5mβ , β = 1, 2, . . . , qm . (3.109)

Let k̃5 (t) = k5 (t) − k5∗ (t). Substituting (3.103), (3.104) and (3.109) into (3.13), we
obtain

ẋ(t) = Ae (x(t) − xm (t)) + Am xm (t) + Bm um (t) + Bm dm (t)

 
∗−1 T T
+ Be K4 K̃1 (t)(x(t) − xm (t)) + K̃2 (t)um (t) + K̃3 xm (t) + k̃5 (t) (3.110)

with e(t) = x(t) − xm (t), we have the tracking error equation

 
ė(t) = Ae e(t) + Be K4∗−1 K̃1T (t)(x(t) − xm (t)) + K̃2 (t)um (t) + K̃3T xm (t) + k̃5 (t) .
(3.111)
Based on (3.111), we choose the adaptive laws as

K̃˙ 50 = K̇50 (t) = −S T BeT P e(t), (3.112)

K̃˙ 5β = K̇5β
T
(t) = −S T BeT P e(t)fβT (t), β = 1, 2, . . . , q. (3.113)

K̃˙ 5mβ = K̇5mβ (t) = −S T BeT P e(t)fmβ

T
(t), β = 1, 2, · · · , p, (3.114)

with P = P T > 0 satisfying AT P + P A = −Q, Q = QT > 0 since Ae is stable.

 45

S ∈ Rp×p satisfies Assumption (A3.8). K1 (t), K2 (t) and K3 (t) are still in (3.95)
-(3.97).
Use the Lyapunov direct method to analysis the stability of the closed-loop system
(3.13) with adaptive controller (3.103) and adaptive laws (3.95) - (3.97) and (3.112)
-(3.114). Choose a positive definite function

V = eT P e + tr[K̃1 Ms−1 K̃1T ] + tr[K̃2T Ms−1 K̃2 ] + tr[K̃3 Ms−1 K̃3T ]

q qm
X X
+ T
tr[k̃5β Ms−1 k̃5β ] + T
tr[k̃5mβ Ms−1 k̃5mβ ] (3.115)
β=1 β=1

as a measurement of the closed-loop error system. Then get the time derivative of
(3.118) which is the negative semidefinite when adaptive laws (3.95) - (3.97) and
(3.112) -(3.114) are substituted.

V̇ = −eT (t)Qe(t) ≤ 0, Q = QT > 0 (3.116)

from (3.116), we can obtain all the desired system properties. 5

Remark 3.2 In fact, there are no essential difference between the disturbance rejec-
tion algorithm applied in Case II and Case III. Actually, Case II is a special form of
Case III with dm (t) = 0. For example in multiple intputs systems, when dm (t) = 0,
it is obviously that k5∗ = k2∗ dm (t) − d(t) = −d(t). The application of this remark is to
be further discussed in Section 4.5. 

3.3.5 Design Based on LDU Parametrization

The adaptive design of previous section for the p-inputs plant (3.11) needs As-
sumption (A3.8): Ms = K4∗ S = (K4∗ S)T = S T K4∗T > 0 for some known matrix
S ∈ Rp×p . This S matrix is analogous to the sign of k4∗ for the case p = 1; however,
 46

the knowledge of such an S matrix is more difficult to obtain then that of the sign
of k4∗ . Relaxation of such knowledge for multivariable adaptive control is important.
In this section, we present an adaptive control scheme for the plant (3.11), using
different knowledge for the gain matrix K4∗ . This adaptive design employs a modified
controller parametrization based on an LDU decomposition of the gain matrix K4∗−1 .
Actually, with LDU decomposition we can address all those six different cases
mentioned in Section 3.1.1 and Section 3.1.2. In this section we will solve Case I with
multi-input systems as an example. The basic principle of adaptive control scheme
and stability analysis for the other two cases are the same.

Gain matrix and design conditions

Proposition 3.1. (LDU decomposition) A matrix Kp ∈ RM ×M with all its leading

principle minors being nonzero has a unique decomposition:

Kp = LD∗ U (3.117)

for some M × M unit(i.e., with all diagonal elements being 1) lower triangular matrix
L and unit upper triangular matrix U , and

 
∗ ∆2 ∆M
D = diag {d∗1 , d∗2 , . . . , d∗M } = diag ∆1 , ,..., . (3.118)
∆1 ∆M −1

This is the well-known LDU decomposition of a nonsingular matrix with nonzero

leading principle minors.
From Proposition 1, we first express the gain matrix K4∗−1 ∈ Rp×p which satisfies
Assumption (A3.8) as
K4∗−1 = LD∗ U (3.119)
 47

for some p × p unit (i.e., with all diagonal elements being 1) lower triangular matrix
L and unit upper triangular matrix U , and

 
∗ ∆2 ∆p
d∗1 , d∗2 , . . . , d∗p

D = diag = diag ∆1 , ,..., (3.120)
∆1 ∆p−1

with ∆i , i = 1, 2, . . . , p, as the leading principle minors of K4∗−1 in (3.68). To use this

decomposition, we assume

(A3.9) All leading principle minors ∆i of the matrix K4∗−1 are nonzero, and their
signs, signs[d∗i ], i = 1, 2, . . . , p, are known;

(A3.10) the matrix L in (3.117) is known.

Controller structure Using (3.68) and (3.117), we express the plant (3.9) as

ẋ(t) = Ae (x(t) − xm (t)) + Am xm (t) + Bm um (t)

+ B(u(t) − K1∗T (x(t) − xm (t)) − K2∗ um (t) − K3∗T xm (t))

= Ae (x(t) − xm (t)) + Am xm (t) + Bm um (t)

+ Be K4∗−1 (u(t) − K1∗T (x(t) − xm (t)) − K2∗ um (t) − K3∗T xm (t))

= Ae (x(t) − xm (t)) + Am xm (t) + Bm um (t)

+ Be LD∗ (U u(t) − U K1∗T (x(t) − xm (t)) − U K2∗ um (t) − U K3∗T xm (t))

= Ae (x(t) − xm (t)) + Am xm (t) + Bm um (t) (3.121)

+ Be LD∗ (u(t) − (I − U )u(t) − U K1∗T (x(t) − xm (t)) − U K2∗ um (t) − U K3∗T xm (t))

where I − U is an upper triangular matrix with zero diagonal elements as U is a unit

upper triangular matrix (whose diagonal elements are 1). From (3.10) and (3.121),
 48

we have the tracking error equation

ė(t) = Ae (x(t) − xm (t)) (3.122)

+ Be LD∗ (u(t) − (I − U )u(t) − U K1∗T (x(t) − xm (t)) − U K2∗ um (t) − U K3∗T xm (t))

Rewrite the tracking error as,

ė(t) = Ae e(t) + Be LD∗ (u(t) − Φ∗0 u(t) − Φ∗T ∗ ∗T

1 (x(t) − xm (t)) − Φ2 um (t) − Φ3 xm (t))

(3.123)
where
Φ∗0 = I − U, Φ∗T ∗T ∗ ∗ ∗T ∗T
1 = U K1 , Φ2 = U K2 , Φ3 = U K3 (3.124)

This error equation motivates the controller structure

u(t) = Φ0 u(t) + ΦT1 (x(t) − xm (t)) + Φ2 um (t) + ΦT3 xm (t) (3.125)

where Φ0 (t), Φ1 (t), Φ2 (t) and Φ3 (t) are the estimates of Φ∗0 , Φ∗1 , Φ∗2 and Φ∗3 , respec-
tively; in particular, the parameter matrix Φ0 has the same special upper triangular
form as that of Φ∗0 = I − U , that is,

 
0 φ12 φ13 · · · φ1p 
 
0 0 φ23 · · · φ2p 
 
 
. . .. .. .. 
Φ0 =  . . (3.126)
. . . . . 

 
 
0 0
 · · · 0 φp−1p 

 
0 ··· ··· 0 0
 49

This special form ensures that the control signals u(t) is implementable from (3.125)
without singularity, that is,

up (t) = [ΦT1 (x(t) − xm (t)) + Φ2 um (t) + ΦT3 xm (t)]p ,

up−1 (t) = φp−1p up (t) + [ΦT1 (x(t) − xm (t)) + Φ2 um (t) + ΦT3 xm (t)]p−1 ,

up−2 (t) = φp−2p−1 up−1 (t) + φp−2p u( t) + [ΦT1 (x(t) − xm (t)) + Φ2 um (t) + ΦT3 xm (t)]p−1 ,
..
.
p
X
u2 (t) = φ2i ui (t) + [Φ0 u(t) + ΦT1 (x(t) − xm (t)) + Φ2 um (t) + ΦT3 xm (t)]2 ,
i=3
p
X
u1 (t) = φ1i ui (t) + [ΦT1 (x(t) − xm (t)) + Φ2 um (t) + ΦT3 xm (t)]1 , (3.127)
i=2

where [v]i denotes the ith row of the vector v.

Adaptive laws To derive at a compact and exact expression of the control law
(3.125) with the special parameter structure (3.126), we let ΦT1i (t) be the ith row of
ΦT1 (t), ΦT2i (t) be the ith row of ΦT2 (t) and ΦT3i (t) be the ith row of ΦT3 (t), i = 1, 2, . . . , p,
and define

θ1 (t) = [φ12 (t), φ13 (t), . . . , φ1p (t), ΦT11 (t), ΦT21 (t), ΦT31 (t)]T ∈ R2n+2p−1

θ2 (t) = [φ23 (t), φ24 (t), . . . , φ2p (t), ΦT12 (t), ΦT22 (t), ΦT32 (t)]T ∈ R2n+2p−2
..
.

θp−2 (t) = [φp−2p−1 (t), φp−2p (t), ΦT1p−2 (t), ΦT2p−2 (t), ΦT3p−2 (t)]T ∈ R2n+p+2

θp−1 (t) = [φp−1p (t), ΦT1p−1 (t), ΦT2p−1 (t), ΦT3p−1 (t)]T ∈ R2n+p+1

θp (t) = [ΦT1p (t), ΦT2p (t), ΦT3p (t)]T ∈ R2n+p (3.128)

 50

which are the estimates of the corresponding θi∗ from the rows Φ∗0 , Φ∗1 , Φ∗2 and Φ∗3 in
(3.124). It follows from (3.123),(3.125) and (3.128) that

 
 θ̃1T (t)ω1 (t)

 
 T
 θ̃2 (t)ω2 (t) 

 
∗
 .. 
ė(t) = Ae e(t) + Be LD  . ,
 (3.129)
 
 T 
θ̃p−1 (t)ωp−1 (t)
 
 
T
θ̃p (t)ωp (t)

where θ̃i (t) = θi (t) − θi∗ , i = 1, 2, . . . , p, and

ω1 (t) = [u2 (t), u3 (t), . . . , up (t), eT (t), rT (t), xTm (t)]T ∈ R2n+2p−1

ω2 (t) = [u3 (t), u4 (t), . . . , up (t), eT (t), rT (t), xTm (t)]T ∈ R2n+2p−2
..
.

ωp−2 (t) = [up−1 (t), up (t), eT (t), rT (t), xTm (t)]T ∈ R2n+p+2

ωp−1 (t) = [up (t), eT (t), rT (t), xTm (t)]T ∈ R2n+p+1

ωp (t) = [eT (t), rT (t), xTm (t)]T ∈ R2n+p (3.130)

we now choose the following adaptive laws for θi (t), i = 1, 2, . . . , p,

θ̇i (t) = −sign[d∗i ]Γi ēi (t)ωi (t), t ≥ 0, (3.131)

where ēi (t) is the ith component of eT (t)P Be L with P = P T > 0, satisfying (3.31),
Γi = ΓTi > 0, and sign[d∗i ] is from Assumption A3.9.
 51

Stability Analysis To analyze the stability properties of the adaptive scheme

(3.131), we consider the positive definite function

p
X
T
V (e, θ̃i , i = 1, 2, . . . , p) = e P e + |d∗i | θ̃iT Γ−1
i (t)θ̃i (3.132)
i=1

The time derivative of V (e, θ̃i ), along the trajectory of (3.131), is

M
X
T
V̇ = 2e (t)P ė(t) + 2 |d∗i | θ̃iT (t)Γ−1
i θ̇i (t)
i=1
XM M
X
T
= 2e (t)P ė(t) + 2 ēi (t)d∗i θ̃iT (t)ωi (t) +2 |d∗i | θ̃iT (t)Γ−1
i θ̇i (t)
i=1 i=1

= −eT (t)Qe(t) (3.133)

Since Q = QT > 0, (3.133) implies that the equilibrium state ec = 0, with ec =

[eT , θ̃1T , . . . , θ̃M
T T
] , of the closed-loop system consisting of (3.129) and (3.131) is uni-
formly stable and its solution ec (t) is uniformly bounded. That is, θi (t), i = 1, 2, . . . , p,
and ė(t) all are bounded. From (3.133), we also have e(t) = x(t) − xm (t) ∈ L2 and so
limt→∞ e(t) = 0 as from the Barbalat lemma. 5

3.4 Simulation Study

To verify the adaptive laws, simulation studies are displayed below. To simplify
the problem and make the result easy to follow, choose n = 2 for all systems below.
Leader parameters are given for simulation study.
In this section, simulation results for all cases including single input systems and
multiple inputs systems are shown. Section 3.4.1 displays simulation results of single
input systems. Section 3.4.2 displays simulation results of multiple inputs systems.
 52

3.4.1 Simulation Study for Single Input Systems

This section shows simulation results of single inputs systems. Simulation results
of three different cases classifying by the disturbance performance in section 3.1.1 are
shown in order. Through the simulation result,we can verify that the adaptive control
laws works well.

Case I: Singe input system without disturbances For numerical study, pa-
rameter matrices in (3.3) and (3.4) are selected as: (n = 2):

       
 1 1 0  1 1 0
Am =   , bm =   A =  ,b =  
−11 −6 8 −1 0 2

and the known Ae and be are chosen as

  
1 1 0
Ae =   be =   .
−9 −4 4

Based on the matching condition (3.8), we can easily obtain the ideal value of
k1∗ , k2∗ and k3∗ .
k1∗T = [−4, −2], k2∗ = 4, /, k3∗T = [−5, −3].

This set of ki (i = 1, 2, 3) is the values that would make up the nominal controller
(3.19). However, this nominal controller maybe unknown in real some experiences.
Therefore, the estimates k1 (t), k2 (t) and k3 (t) of k1∗ , k2∗ and k3∗ will be determined from
adaptive laws and used in the adaptive controller (3.22).
For the purpose of the simulation, the ideal gains were calculated. They are
used to show how a fixed gain controller cannot handle the parameter unknown case.
Since k1 (0), k2 (0) and k3 (0) can be chosen arbitrary, here we choose ki (0) = 0.5ki∗ , (i =
 53

1, 2, 3). r = sin(t) + cos(t). The initial leader and follower state are xm = [0, 0]T and
x = [0, 0]T , respectively.
The simulation result with a fixed gain controller for the state tracking of x is
shown in figure 3.1. Likewise, the simulation result with adaptive controller for x is
shown in figure 3.2. Tracking errors in both fixed gain control and adaptive control
are shown in figure 3.3. Parameter errors ki (t) − ki∗ (t) are shown in figure 3.4.

Case II: Single input systems with disturbance acting on follower systems
For numerical study, parameter matrices in (3.5) and (3.4) are selected as: (n = 2):

       
 1 1 0  1 1 0
Am =   , bm =   , A =  , b =  
−11 −6 8 −1 0 2

The disturbance (3.6) is

d(t) = −5 − 4 sin(5t).

Ae , be are chosen as    
1 1 0
Ae =   , be =   ,
−9 −4 4

Parameter matrices/vectors A, b, Am , b,Ae and be adopted for simulation in Case

I and Case II are the same. It is obviously that ki∗ , i = 1, 2, 3 are in these two
special cases are the same. According to the value of d(t). We can conclude that
k5∗ = 5 + 4 sin(5t).
For the purpose of the simulation, the ideal gains were calculated. They are used
to show how a fixed gain controller cannot handle the parameter unknown case and in
this case, particularly, an unknown external bounded disturbance. Since k1 (0), k2 (0)
 54

−1

−2

−3
0 5 10 15 20 25 30
Follower state x1 and leader state xm1 vs. time(sec)

−1

−2

−3
0 5 10 15 20 25 30
Follower state x2 and leader state xm2 vs. time(sec)

Figure 3.1: Follower state (solid) vs. leader state (dotted) with fixed gain control.

−1

−2
0 5 10 15 20 25 30
Follower state x1 and leader state xm1 vs. time(sec)

−1

−2

−3
0 5 10 15 20 25 30
Follower state x2 and leader state xm2 vs. time(sec)

Figure 3.2: Follower state (solid) vs. leader state (dotted) with adaptive gain control.
 55

0.6
e1(t)
e2(t)
0.4

0.2

−0.2

−0.4
0 5 10 15 20 25 30
Tracking error e(t) with adaptive control vs. time(sec)

0.8
e (t)
1
0.6 e2(t)
0.4

0.2

−0.2

−0.4

−0.6
0 5 10 15 20 25 30
Tracking error e(t) with fixed gain control vs. time(sec)

Figure 3.3: Tracking errors with adaptive control vs. fixed gain control.

2
K (t)
1,1
K (t)
1 1,2

−1
0 5 10 15 20 25 30
Parameter errors k − k* vs. time(sec)
1,i 1,i

−1.8
K2(t)
−1.9

−2

−2.1

−2.2
0 5 10 15 20 25 30
Parameter error k2 − k*2 vs. time(sec)

3
K3,1(t)
2.5 K3,2(t)

1.5

1
0 5 10 15 20 25 30
Parameter errors k3,i − k*3,i vs. time(sec)

Figure 3.4: Parameter errors ki (t) − ki∗ with adaptive control (i = 1, 2, 3).
 56

and k3 (0) can be chosen arbitrary, here we choose ki (0) = 0.8ki∗ , (i = 1, 2, 3). k50 (0) =
5 × 0.8 = 4, k5β (0) = 4 × 0.8 = 3.6. r = sin(t) + cos(t). The initial leader and follower
state are xm = [1, 0]T and x = [−1, 5]T , respectively.
The simulation result with a fixed gain controller for the state tracking of x is
shown in figure 3.5. Likewise, the simulation result with an adaptive controller for
the state tracking of x is shown in figure 3.6. Tracking errors in both fixed gain
control and adaptive control are shown in figure 3.7.

Case III: Single input system with disturbance acting on leader and fol-
lower systems For numerical study, parameter matrices in (3.5) and (3.7) are
selected as: (n = 2):

       
 1 1 0  1 1 0
Am =   , bm =   A =  ,b =  ,
−11 −6 8 −1 0 2

Disturbances acting on (3.5) and (3.7) are chosen as follows:

d(t) = −5 + (−4) sin(5t),

dm (t) = 4 + 3.6 sin(5t).

Ae and be are chosen as

  
1 1 0
Ae =  , b =
 e  
−9 −4 4

Parameter matrices/vectors A, b, Am and b adopted for simulation in Case I and

Case II are the same.So the ideal controller parameters ki∗ , i = 1, 2, 3 are the same
as the ideal controller parameters in Case I and Case II. According to (3.52), k5∗ =
 57

−1

−2

−3
0 2 4 6 8 10 12 14 16 18 20
Follower state x1 and leader state xm1 vs. time(sec)

−1

−2

−3
0 2 4 6 8 10 12 14 16 18 20
Follower state x2 and leader state xm2 vs. time(sec)

Figure 3.5: Follower state (solid) vs. leader state (dotted) without disturbance rejec-
tion.

1.5

0.5

−0.5

−1

−1.5

−2
0 2 4 6 8 10 12 14 16 18 20
Follower state x1 and leader state xm1 vs. time(sec)

−1

−2

−3
0 2 4 6 8 10 12 14 16 18 20
Follower state x2 and leader state xm2 vs. time(sec)

Figure 3.6: Follower state (solid) vs. leader state (dotted) with adaptive disturbance
rejection.
 58

5
e1(t)
4 e2(t)
3

−1

−2
0 2 4 6 8 10 12 14 16 18 20
Tracking error e with adaptive conrol vs. time(sec)

5
e1(t)
4 e2(t)
3

−1

−2
0 2 4 6 8 10 12 14 16 18 20
Tracking error e with fixed gain conrol vs. time(sec)

Figure 3.7: Tracking errors with adaptive disturbance rejection vs. without distur-
bance rejection.

21 + 16 sin(5t) − 3.6 sin(5t).

For the purpose of the simulation, the ideal gains were calculated. They are used to
show how a fixed gain controller cannot handle the parameter unknown case. Since
k1 (0), k2 (0) and k3 (0) can be chosen arbitrary, here we choose ki (0) = 0.8ki∗ , (i =
1, 2, 3). k50 (0) = 21 = 16.8, k5mβ (0) = 16 × 8 = 12.8, k5β (0) = 3.6 × 8 = 2.88.
r = sin(t). The initial leader and follower state are xm = [4, 5]T and x = [3, 1]T ,
respectively.
The simulation result with a fixed gain controller for the state tracking of x is
shown in figure 3.8. Likewise, the simulation result with an adaptive controller for
the state tracking of x is shown in figure 3.9. Tracking errors in both fixed gain
control and adaptive control are shown in figure 3.10.
 59

10

2
0 5 10 15 20 25 30
Follower state x1 and leader state xm1 vs. time(sec)

−5

−10

−15
0 5 10 15 20 25 30
Follower state x and leader state x vs. time(sec)
2 m2

Figure 3.8: Follower state (solid) vs. leader state (dotted) without disturbance rejec-
tion

10

3
0 5 10 15 20 25 30
Follower state x1 and leader state xm1 vs. time(sec)

10

−5

−10

−15
0 5 10 15 20 25 30
Follower state x2 and leader state xm2 vs. time(sec)

Figure 3.9: Follower state (solid) vs. leader state (dotted) with adaptive disturbance
rejection
 60

6
e1(t)
e2(t)
4

−2

−4
0 5 10 15 20 25 30
Tracking error e with an adaptive control vs. time(sec)

4
e1(t)
3
e2(t)
2

−1

−2

−3

−4
0 5 10 15 20 25 30
Tracking error e with fixed gain control vs. time(sec)

Figure 3.10: Tracking errors with adaptive disturbance rejection vs. without distur-
bance rejection.

3.4.2 Simulation Study with Multiple Inputs

This section shows simulation results of single inputs systems. Simulation results
of three different cases classifying by the disturbance performance in section 3.2.1 are
shown in order. Through the simulation result,we can verify that the adaptive control
laws works well.

Case I: Multiple inputs systems without disturbance For numerical study,

parameter matrices in (3.11) and (3.12) are selected as: (n = 2, p = 2, m = 3):

       
 1 1 2 0 2  1 1 2 0
Am =   , Bm =  ,A =  ,B =  .
−11 −6 0 2 0 −1 0 0 2
 61

The known parameter Ae and Be are

   
1 1 6 0
Ae =   , Be =  .
−9 −4 0 6

Based on the matching condition (3.68), we can easily obtain the ideal value of
K1∗ , K2∗ and K3∗ .

     
0 0 1 0 1 0 0
K1∗T = ∗
 , K2 = 
∗T
 , K3 =  .
−4 −2 0 1 0 −5 −3

For the purpose of the simulation, the ideal gains were calculated. They are used
to show how a fixed gain controller cannot handle the parameter unknown case. Since
K1 (0), K2 (0) and K3 (0) can be chosen arbitrary, here we choose Ki (0) = 0.5Ki∗ , (i =
1, 2, 3). r = [sin(t), cos(t), sin(t) + cos(t)]T . The initial leader and follower state are
xm = [0, 0]T and x = [0, 0]T , respectively.
The simulation result with a fixed gain controller for the state tracking of x is
shown in figure 3.11. Likewise, the simulation result with adaptive controller for x is
shown in figure 3.12. Tracking errors in both fixed gain control and adaptive control
are shown in figure 3.13. Parameter errors Ki (t) − Ki∗ (t) are shown in figure 3.14.

Case II: Multiple inputs systems with disturbance acting on follower sys-
tems For numerical study, parameter matrices in (3.13) and (3.15) are selected as:
 62

−5
0 5 10 15 20 25 30 35 40
Follower state x1 and leader state xm1 vs. time(sec)

10

−5

−10
0 5 10 15 20 25 30 35 40
Follower state x and leader state x vs. time(sec)
2 m2

Figure 3.11: Follower state (solid) vs. leader state (dotted) with fixed gain control.

−5
0 5 10 15 20 25 30 35 40
Follower state x1 and leader state xm1 vs. time(sec)

10

−5

−10
0 5 10 15 20 25 30 35 40
Follower state x2 and leader state xm2 vs. time(sec)

Figure 3.12: Follower state (solid) vs. leader state (dotted) with adaptive gain control.
 63

2
e1(t)
1.5 e2(t)

0.5

−0.5

−1
0 5 10 15 20 25 30 35 40
Tracking error e with adaptive control vs. time(sec)

2
e1(t)

1 e2(t)

−1

−2

−3
0 5 10 15 20 25 30 35 40
Tracking error w with fixed gain control vs. time(sec)

Figure 3.13: Tracking errors with adaptive control vs. fixed gain control.

−5

−10

−15
0 20 40 60 80 100 120 140 160 180 200
Parameter error K − K* vs. time(sec)
1 1

−2

−4
0 20 40 60 80 100 120 140 160 180 200
Parameter error K2 − K*2 vs. time(sec)

−5

−10

−15
0 20 40 60 80 100 120 140 160 180 200
Parameter error K3 − K*3 vs. time(sec)

Figure 3.14: Parameter errors Ki (t) − Ki∗ with adaptive control (i = 1, 2, 3).
 64

(n = 2):

      
0 1 1 −2 −2 3  0 1
Am =   , Bm =  ,A =  ,B =  ,
−1 −3 4 1 1 −2 1 2

The disturbance acting on the follower is

d(t) = [−5, −5]T + [1, 2]T sin(5t)

Ae and Be are
   
1 1 1 −2
Ae =   , Be =  .
−9 −4 4 1

Based on the matching condition (3.68), we can easily obtain the ideal value of
K1∗ , K2∗ and K3∗ .

     
−16 2  2 5  −6 3 
K1∗T = ∗
 , K2 = 
∗T
 , K3 =  ,
3 −2 1 −2 2 −2

∗
According to the equation k5β = −d(t), k5∗ = [5, 5]T + [−1, −2]T sin(5t).
For the purpose of the simulation, the ideal gains were calculated. They are used
to show how a fixed gain controller cannot handle the parameter unknown case. Since
K1 (0), K2 (0) and K3 (0) can be chosen arbitrary, here we choose Ki (0) = 0.9ki∗ , (i =
∗ ∗
1, 2, 3). k50 = [−4.5, −4.5]T , k5β = [−0.8, 1.8]T . r = [sin(t), cos(t)]T . The initial
leader and follower state are xm = [1, 1]T and x = [2, 3]T , respectively.
The simulation result without adaptive disturbance rejection for the state tracking
of x is shown in figure 3.15. Likewise, the simulation result with adaptive disturbance
 65

rejection for x is shown in figure 3.16. Tracking errors corresponding with disturbance
rejection and non-disturbance rejection are shown in figure 3.17.
Remark 3.3 To show the effectiveness of the adaptive compensator, figure 3.15 is the
result making by an adaptive controller without an adaptive disturbance compensator
term, i.e., we adopt the adaptive controller (3.72) rather than (3.91). Figure 3.16 is
created by the adaptive disturbance compensator (3.91). 

Case III: Multiple inputs systems with disturbance acting on leader and
follower systems For numerical study, parameter matrices in (3.13) and (3.16) are
selected as: (n = 2):

  
    
0 1 1 −2 −2 3  0 1
Am =   , Bm =  ,A =  ,B =  ,
−1 −3 4 1 1 −2 1 2

The disturbance (3.14) and (3.17) are

dm (t) = [−5, −5]T + [1, 2]T sin(5t)

d( t) = [4, 4]T + [2, 1]T cos(5t)

Ae and Be are
   
1 1 1 −2
Ae =  , B
 e = .
−9 −4 4 1

Based on the matching condition (3.45), we can easily obtain the ideal value of
 66

−2

−4

−6
0 10 20 30 40 50 60 70 80 90 100
Follower state x1 and leader state xm1 vs. time(sec)

−1

−2

−3
0 10 20 30 40 50 60 70 80 90 100
Follower state x and leader state x vs. time(sec)
2 m2

Figure 3.15: Follower state (solid) vs. leader state (dotted) without disturbance
rejection.

−2

−4

−6
0 10 20 30 40 50 60 70 80 90 100
Follower state x1 and leader state xm1 vs. time(sec)

−1

−2

−3
0 10 20 30 40 50 60 70 80 90 100
Follower state x2 and leader state xm2 vs. time(sec)

Figure 3.16: Follower state (solid) vs. leader state (dotted) with adaptive disturbance
rejection.
 67

1.5
e1(t)
e2(t)
1

0.5

−0.5
0 10 20 30 40 50 60 70 80 90 100
Tracking error e with adaptive disturbance rejection vs. time(sec)

1
e1(t)
0.8 e2(t)
0.6

0.4

0.2

−0.2

−0.4
0 10 20 30 40 50 60 70 80 90 100
Tracking error e without adaptive disturbance rejection vs. time(sec)

Figure 3.17: Tracking errors with adaptive disturbance rejection vs. without distur-
bance rejection.

K1∗ , K2∗ and K3∗ .

     
−16 2  2 5  −6 3 
K1∗T = ∗
 , K2 = 
∗T
 , K3 =  .
3 −2 1 −2 2 −2

Parameter matrices A, B, Am and Bm adopted for simulation in Case I and Case

II are the same. It is obviously that ki∗ , i = 1, 2, 3 are in these two special cases are
the same. k5∗ = [23, −9]T + [28, −4]T sin(5t) − [1, 2]T cos(5t).
For the purpose of the simulation, the ideal gains were calculated. They are used
to show how a fixed gain controller cannot handle the parameter unknown case. Since
K1 (0), K2 (0) and K3 (0) can be chosen arbitrary, here we choose Ki (0) = 0.9ki∗ , (i =
∗ ∗
1, 2, 3). k50 = 4, k5j = −3.6. r = [sin(t), cos(t)]T . The initial leader and follower state
are xm = [1, 1]T and x = [2, 3]T , respectively.
 68

The simulation result without adaptive disturbance rejection for the state track-
ing of x is shown in figure 3.18 (i.e., created by (3.72)). Likewise, the simulation
result with adaptive disturbance rejection for x is shown in figure 3.19 (i.e., cre-
ated by (3.107)). Tracking errors corresponding with disturbance rejection and non-
disturbance rejection are shown in figure 3.20.

Simulation study with adaptive control based on LDU decomposition The

following simulation results show the tracking performance without disturbances
which is exactly the case we discussed in Section 3.3.5.
For numerical study, parameter matrices in (3.6) and (3.7) are selected as: (n =
2, p = 2):

       
 1 1 2 0  1 1 2 0
Am =   , Bm =  ,A =  ,B =  ,
−11 6 0 2 −1 0 0 2

Ae and Be are    
1 1 1 −2
Ae =   , Be =  .
−9 −4 4 1

Based on the matching condition (3.68), we can easily obtain the ideal value of
Φ∗0 , Φ∗1 , Φ∗2 and Φ∗3 .

       
0 −2 ∗T −8 −4 ∗ 1 2 ∗T −10 −6
Φ∗0 =   Φ1 =   Φ2 =   Φ3 =  
0 0 −4 −2 0 1 −5 −3

For the purpose of the simulation, the ideal gains were calculated. They are used
to show how a fixed gain controller cannot handle the parameter unknown case. Since
θ1 (0), θ2 (0) and θ3 (0) can be chosen arbitrary, here we choose θi (0) = 0.8θi∗ , (i = 1, 2).
r = [sin(t) cos(t)]T . The initial leader and follower state are xm = [2.5, 1]T and
 69

−2

−4

−6

−8
0 5 10 15 20 25 30
Follower state x1 and leader state xm1 vs. time(sec)

12

10

0
0 5 10 15 20 25 30
Follower state x and leader state x vs. time(sec)
2 m2

Figure 3.18: Follower state (solid) vs. leader state (dotted) without disturbance
rejection.

−2

−4

−6

−8
0 5 10 15 20 25 30
Follower state x1 and leader state xm1 vs. time(sec)

12

10

0
0 5 10 15 20 25 30
Follower state x2 and leader state xm2 vs. time(sec)

Figure 3.19: Follower state (solid) vs. leader state (dotted) with adaptive disturbance
rejection.
 70

4
e1(t)
2 e2(t)

−2

−4

−6

−8
0 5 10 15 20 25 30
Tracking error e with adaptive disturbance rejection vs. time(sec)

6
e1(t)
4 e2(t)

−2

−4
0 5 10 15 20 25 30
Tracking error e without adaptive disturbance rejection vs. time(sec)

Figure 3.20: Tracking errors with adaptive disturbance rejection vs. without distur-
bance rejection.

x = [0.5, 2]T , respectively.

The simulation result with a fixed gain controller for the state tracking of x is
shown in figure 3.21. Likewise, the simulation result with adaptive controller for x is
shown in figure 3.22. Tracking errors in both fixed gain control and adaptive control
are shown in figure 3.23.

3.5 Summary

In this chapter basic adaptive control theory relevant to this research is intro-
duced. Since this is a state tracking problem, in multi-agent system we also called
this kind of problem as ”leader-following consensus” problem. Leader-following con-
sensus problems presented in this chapter are categorized into six cases in total in-
cluding three single input cases and three multiple inputs cases. Stability analyses are
 71

−1

−2
0 5 10 15 20 25
Follower state x1 and leader state xm1 vs. time(sec)

−2

−4

−6
0 5 10 15 20 25
Follower state x and leader state x vs. time(sec)
2 m2

Figure 3.21: Follower state (solid) vs. leader state (dotted) with adaptive gain control
(LDU decomposition).

−1

−2
0 5 10 15 20 25
Follower state x1 and leader state xm1 vs. time(sec)

−2

−4

−6
0 5 10 15 20 25
Follower state x2 and leader state xm2 vs. time(sec)

Figure 3.22: Follower state (solid) vs. leader state (dotted) with adaptive gain control
(LDU decomposition).
 72

2
e1(t)
e2(t)
1

−1

−2

−3
0 5 10 15 20 25
Tracking error e with adaptive control vs. time(sec)

6
e1(t)
e2(t)
4

−2

−4
0 5 10 15 20 25
Tracking error e with fixed gain control vs. time(sec)

Figure 3.23: Tracking errors with adaptive control vs. fixed gain control (LDU de-
composition).

given for each case to show with the adaptive controllers and the corresponding adap-
tive laws, closed-loop systems in all six cases have obtained the desired properties.
Also, in the last subsection in Section 3.3 which focuses on the multiple input sys-
tems, LDU decomposition is applied. LDU decomposition is used because in practice
some parameters are hard to get, so we need another functional adaptive controller
structure with different adaptive laws. Remind that with LDU decomposition, both
disturbances-free cases and cases with disturbances can be solved. However, we just
show the disturbance-free case to show the basic principle. Many other literatures
show more details which are listed in Section 3.3.5. In Section 3.4, simulation results
are shown in order to verify that the adaptive controllers with corresponding adaptive
laws work well. Simulation results show the difference between adaptive control and
fixed gain control directly.
Chapter 4

Adaptive Leader-Following
Consensus for Multiple Agents

In Chapter 3, we discussed about one leader-one follower consensus adaptive con-

trol strategy, from which we derive a new control strategy can be applied on multi-
agent systems with directed graph. In Section 4.1, system dynamic expressions are
given and the control objective is discussed. Algebraic graph theory which is used
to represent the interactions among multi-agent systems is to be discussed in Section
4.2. In Section 4.3, a distributed adaptive control scheme for multi-agent system
describing by direct graphs is developed. Disturbances are introduced in Section 4.4,
and adaptive control compensator term will be added and the corresponding adaptive
control scheme will be discussed in the same section. In the last section, simulation re-
sults are shown to indicate the capability of the distributed adaptive control schemes
presented in this chapter.

73
 74

4.1 Problem Statement

Consider a set of N unknown follower agents. The dynamic system of the ith
follow agent can be expressed as

ẋi (t) = Ai xi (t) + Bi ui (t), i = 1, . . . , N (4.1)

where xi (t) ∈ Rn is the state of the ith agent, ui (t) ∈ Rpi is the control input(in
multi-agent system, we also call it control protocol), All parameter matrices Ai ∈
Rn×n , Bi ∈ Rn×pi , i = 1, . . . , N in (4.1) are unknown. The dynamics of the leader is
given by

ẋ0 = A0 x0 + B0 u0 (t) (4.2)

where x0 ∈ Rn is the state of the leader, u0 (t) ∈ Rm is the bounded reference signal
and A0 ∈ Rn×n , B0 ∈ Rn×m are unknown constant matrices.
The control objective is to design a distributed control protocol using local infor-
mation for each follower to drive all follower consensus on their states with a given
leader, i.e., make all the followers track the give leader on states eventually. Such a
control problem is defined as leader-following consensus problem. The control objec-
tive can be described mathematically as

lim (xi (t) − x0 (t)) = 0, i = 1, . . . , N. (4.3)

t→∞

The above equation means if the control objective is achieved, all states errors between
follower agents and the leader should go to zero as t goes to infinity.
 75

4.2 Algebraic Graph Theory

The interaction graphs among N + 1 agents including one leader and N followers
are denoted by a directed graph G = (V, E) with a set of vertices V and a set of
directed edges (ordered pairs of vertices) E ⊆ V × V. A vertex represents an agent
and the directed edge (vj , vi ) denotes that agent vi can obtain the information from
agent vj including both the state information and the input information, but not vice
versa. In this case, we say that vj is one of the neighbors of vi (although the link
between them is directed). Define a neighborhood set Ni for the follower agent vi
such that Ni = {vj ∈ V : (vj , vi ) ∈ E} for i = 1, . . . , N . v0 denotes the leader in the
multi-agent system. It also could be a neighbor of the follower agent vi if vi can get
the information from v0 (i.e. (v0 , vi ) ∈ E). A directed path is a sequence of ordered
edges of the form (vi1 , vi2 ), (vi2 , vi3 ), . . . in a directed graph, where vij ∈ V. Define
an indegree (the number of head endpoints adjacent to a vertex) matrix D such that
D = diag{n1 , n2 , . . . , nN } = diag{deg(v1 ), deg(v2 ), . . . , deg(vN )} where deg(·) denotes
the indegree of an agent. This indegree matrix D represents the total number of the
neighbors of the follower agent vi , i.e., the total agent number in the neighborhood
set Ni from i = 1 to N .
To describe the relationship between the leader and followers, we denote B =
diag {µi } if the agent i can get the information from the leader directly then µi = 1
otherwise µi = 0. Denote Q = L + B. matrix Q is a matrix which can fully described
the N + 1 order graph including N followers and one leader. It is obviously that
elements qii on diagonal are the total number of agents in the neighbor sets Ni of ith
agent respectively. In order to make the equation more clear, we denote ni equals
qii to represent the total agent number in neighbor set Ni . More examples show in
section 4.5.
 76

In order to make the leader-following consensus realized, we give two assumptions

on the interaction graphs.

Assumption 4.1. For each agent vi , there should exist at least one directed path
(v0 , v1 ), (v1 , v2 ), . . . , (vi−1 , vi ) which starts from the leader and ends at the agent vi .

Assumption 4.2. Directed graph G is simple, which means that it has no loops, and
no multiple arcs (arcs with same starting and ending nodes).

Remark 4.1 Many papers focusing on multi-agent leader-following consensus prob-

lem demonstrate that the controller they proposed work perfectly for complex multi-
agent systems which have many different kinds of edges. But in practical, engineers
always consider how to achieve the objective more effectively. From this considera-
tion, we find out the simplest multi-agent system structure which our proposed control
protocol can still work with. Being derived from Assumption 4.1 and Assumption 4.2,
(4.4) is the corresponding matrix Q of such the simplest multi-agent systems after
elementary transformations.

 
1 0
 
 
∗ 1 
 

 .
 (4.4)
 .
.. .. 

 
 
∗ ∗ 1
N ×N

with * representing −1 or 0. (4.4) means for a multi-agent system including N

follower agents and one leader, as long as there exist N edges and each follower agent
has in-degree one, the distributed adaptive control protocol proposed in this chapter
is applicable. 
 77

4.3 Adaptive Control Design

To make adaptive control design work, several design conditions should be sat-
isfied. Design conditions are presented in Section 4.3.1 and a distributed adaptive
control protocol with adaptive laws is presented in Section 4.3.2.

4.3.1 Design Conditions

To meet the control objective and solve the leader-following consensus problem,
we need several design conditions to be satisfied.

(A4.1) all the eigenvalues of A0 are in the open left-half complex plane;

(A4.2) u0 (t) is bounded and piecewise continuous;

∗ ∗
(A4.3) there exist two matrices K1i ∈ Rn×pi and K4i ∈ Rpi ×pi for each follower agent
vi , which satisfy the following equations

∗T ∗
Ae = Ai + Bi K1i , Be = Bi K4i . (4.5)

where Ae ∈ Rn×n is a stable and known matrix, such that ATe P + P Ae = −Q <
0, Q = QT > 0 is satisfied, and Be ∈ Rn×p is also a known matrix.

∗
(A4.4) there exist two matrices K2i0 ∈ Rpi ×m , K3i0
∗
∈ Rn×pi such that

∗T ∗
A0 = Ai + Bi K3i0 , B0 = Bi K2i0 (4.6)

are satisfied, if the leader is one of the neighbors of the follower agent vi (i.e.,
 78

(v0 , vi ) ∈ E). Otherwise, if (v0 , vi ) ∈

/ E,

∗T ∗
Aj = Ai + Bi K3ij , Bj = Bi K2ij (4.7)

∗ ∗
should be satisfied for some K3ij ∈ Rn×pi and K2ij ∈ Rpi ×pj for each pair of
(vj , vi ) ∈ E (j 6= 0).

(A4.5) there is a known matrix Si ∈ Rpi ×pi for each follower i such that K4i
∗
Si is
∗ ∗ ∗T
symmetric and positive definite: Ms = K4i Si = (K4i Si )T = SiT K4i > 0.

Assumptions (A4.1) and (A4.2) are for a stable, well-defined reference system
(A0 , B0 ) with a bounded output y0 (t). Assumptions (A4.3) and (A4.4) are called the
matching conditions. Assumption (A4.4) indicates that we classify the N followers
into two groups. In the first group, each follower has a direct access to obtain the
∗
information of the leader, thus for each pair of (v0 , vi ), there exist two matrices K2i0
∗
and K3i0 satisfying (4.6). Followers in the other group have no direct accesses to the
∗ ∗
leader, thus there exist two matrices K2ij and K3ij for each pair of (vj , vi ), j 6= 0. In
summary, for each directed edges (ordered pair) (vj , vi ) in the multi-agent system, no
∗ ∗
matter vj is the leader or a follower, there exist a set of corresponding K2ij and K3ij
(0 ≤ j ≤ N ).
Suppose all the parameters Ai , Bi for i = 0, 1, . . . , N are known and Assumptions
(4.5) - (4.7) satisfied, an ideal distributed adaptive protocol which can achieve the
control objective is designed as

1 X
u∗i (t) = ∗T ∗ ∗T


K1i (xi (t) − xj (t)) + K2ij uj (t) + K3ij xj (t) (4.8)
ni v ∈N
j i

with ni being the total agent number in the neighbor set Ni .

 79

We substitute (4.8) into (4.1), the ith closed-loop subsystem (4.1) becomes

1 X
ẋi (t) = Ai xi (t) + Bi ∗
(K ∗T (xi (t) − xj (t)) + K2ij ∗T
uj (t) + K3ij xj (t)) (4.9)
ni v ∈N 1ij
j i

1 X
= (Ae (xi (t) − xj (t)) + Aj xj (t) + Bj uj (t))
ni v ∈N
j i

To begin with a control design, define a local tracking error for each agent here in
this multi-agent case as follow

1 X
ei (t) = xi (t) − xj (t). (4.10)
ni v ∈N
j i

for i = 1, . . . , N , as a measurement to show the disagreement between the follower

i and the average of its neighbors on the states. The motivation of defining such a
local state tracking error is shown in the following lemma.

Lemma 4.1. Under Assumptions (A4.1) and (A4,2), if limt→∞ ei (t) = 0 holds, then
limt→∞ (xi (t) − x0 (t)) = 0 holds for all i = 1, . . . , N .

Proof : First we use a one leader and two followers case to verify the Lemma 4.1. For
a one leader two followers directed graph, there are four formation possible cases in
total.
Case I: In figure 4.1, since we have known that e1 (t) = x1 (t) − x0 (t) → 0 and
e2 (t) = x2 (t) − x1 (t) as t → ∞, then obviously we can get x1 (t) − x0 (t) → 0 and
x2 (t) − x0 (t) → 0 as t → ∞.

0 1 2
Figure 4.1: Case I: Interaction graph of two followers and one leader
aaaaa
aaaaaaaa
 80

1 2
Figure 4.2: Case II: Interaction graph of two followers and one leader

Case II: In figure 4.2, we have known that e1 (t) = x1 (t) − x0 (t) → 0 and e2 (t) =
x2 (t)−x0 (t) → 0 as t → ∞, then obviously we can get x1 (t) → x0 (t) and x2 (t) → x0 (t)
as t → ∞.
Case III: In figure 4.3, since we have known that e1 (t) = x1 (t) − 21 x0 (t) − 12 x2 (t) → 0

1 2
Figure 4.3: Case III: Interaction graph of two followers and one leader

and e2 (t) = x2 (t) − x0 (t) as t → ∞, then obviously we can get x1 (t) − x0 (t) → 0 and
x2 (t) − x0 (t) → 0 as t → ∞.
Case IV: Case IV is similar with Case III. We can also verify Lemma 4.1 quickly.

1 2
Figure 4.4: Case IV: Interaction graph of two followers and one leader

After the verification, we will start our proof. Under Assumption 4.1, for each follower,
there exists at least one directed path from the leader to the follower, which indicates
that each directed graph with N followers and one leader in this thesis consists of at
least one basic structure which is like figure 4.5. In figure 4.5, j = 1, . . . , N .

0 1 2 ··· j-1 j

Figure 4.5: A basic structure exists in directed graphs

 81

We assume that for all followers, limt→∞ ei (t) = 0, i = 1, . . . , j, j < N in figure 4.5.
Each directed graph may has several basic structures with different orders (i.e., j can
be different). Under the condition that limt→∞ ei (t) = 0 holds for i = 1, . . . , j, j < N ,
if we can prove that all followers in figure 4.5 can achieve the global tracking, i.e.,
limt→∞ (xi (t)−x0 (t)) = 0 no matter how the basic structure changes, we can conclude
that the Lemma 4.1 is true.
We have two different methods to make figure 4.5 more complex. The first method
is to add neighbors on follower v1 in figure 4.5 (i.e., the follower who can get the
information from the leader directly). The second method is to add neighbors on
follower v2 to vj (i.e., followers who can not get information from the leader directly).
Before we making changes on figure 4.5, we assume that all followers in figure 4.5
have already achieved the global tracking which means limt→∞ (xi (t) − x0 (t)) = 0, i =
1, . . . , j.

First method: there is only one possibility to add neighbors on follower v1 un-
der Assumption 4.1 and Assumption 4.2, which is to add a new follower outside into
figure 4.5 (maybe a follower from other basic structures in the same directed graph).
Figure 4.6 shows the interaction graph when we add a new follower vj+1 into figure
4.5. According to the condition of Lemma 4.1, we have ej+1 (t) = xj+1 (t) − x0 (t) → 0

0 1 2 ··· j-1 j

j+1

Figure 4.6: Interaction graph after adding one follower on follower v1

as t → ∞ and e1 = x1 − 12 x0 (t) − 12 xj+1 (t). So we have x1 (t) − x0 (t) → 0 as t → ∞.

Since follower v2 to vj do not change, it is obviously that limt→∞ (xi (t) − x0 (t)) = 0. If
 82

there are more neighbors adding on follower vi , the situations will be the same. aaaa
aaaa
Second method: Add followers on the follower who does not have the access to get
the information from the leader directedly. aaaa
Case I: Add the leader as a neighbor of follower vk (k 6= 0 and k < N , see figure
4.7). Here we use mathematic induction to prove that the lemma is true under Case
I by the second method.

0 1 2 ··· k ··· j-1 j

Figure 4.7: Interaction graph after adding one follower on follower v1

First, when adding the leader as a neighbor of follower vk , since all limt→∞ ei (t) =
0, we have limt→∞ ek (t) = limt→∞ (xk (t) − 12 x0 (t) − 12 xk−1 (t)) = 0. Since we as-
sume that all the followers have already achieved the global tracking before we start
adding neighbors, we have limt→∞ (xk−1 (t) − x0 (t)) = 0. Thus, obviously we have
limt→∞ (xk (t) − x0 (t)) = 0.
Then we assume when we adding n followers on the follower vk including the
leader, we can get

1 X
lim ei (t) = lim (xk (t) − xk−1 (t)) = 0
t→∞ t→∞ n v ∈N
k−1 kn

according to limt→∞ (xi (t) − x0 (t)) = 0. Nkm denotes the neighbors of vk when it has
m neighbors (m¡N).
 83

Under the situation when adding n followers on follower vk , we add the (n + 1)th
follower va on the follower vk including the leader, we have

 
1 X
lim ek (t) = lim xk (t) − ( xk−1 (t) + xa (t)) = 0
t→∞ t→∞ n + 1 v ∈N
k−1 m+1

Since all the other followers do not change, so xa → x0 as t → ∞, xk−1 → x0 as

t → ∞, xk−1 ∈ Nkn+1 (Nkn+1 represents the neighborhood when follower vk has n
neighbors). Thus, when there are n + 1 followers on followers vk , according to

1 X
lim ek (t) = lim (xk (t) − ( xk−1 (t) + xa (t))) = 0
t→∞ t→∞ nk+1 vk−1 ∈Nk

lim xk−1 (t) = 0, lim xa (t) = 0

t→∞ t→∞

we have,
lim (xk (t) − x0 (t)) = 0
t→∞

Then,
lim (xi (t) − x0 (t)) = 0, i = 1, . . . , j
t→∞

Case II: Add other followers as neighbors of follower vk (k 6= 0 and k < N , see
figure 4.8).

0 1 2 ··· k ··· j-1 j

Figure 4.8: Interaction graph after adding one follower for follower vk

We still use mathematic induction to prove that the lemma is true under Case II
by the second method.
 84

First, when adding follower xa , (a 6= 0), as a neighbor of follower vk , since all

limt→∞ ei (t) = 0, we have limt→∞ ek (t) = limt→∞ (xk (t)− 21 xa (t)− 12 xk−1 (t)) = 0. Since
we assume that all the followers have already achieved the global tracking before we
start adding neighbors, we have limt→∞ (xk−1 (t) − x0 (t)) = limt→∞ (xa (t) − x0 (t)) = 0.
Thus we have limt→∞ (xk (t) − x0 (t)) = 0.
When adding n followers or n + 1 followers into figure 4.5 under Case II, the
situation will be similar with Case I. Then,

lim (xi (t) − x0 (t)) = 0, i = 1, . . . , j

t→∞

In summary, we have proved that the global tracking can be achieved if we known
that all limt→∞ ei (t) = 0 whatever changes we make on a basic structure in directed
graph. Also the directed graphs under Assumption (A4.1) and (A4.2) consist of
several different basic structures. Thus we can prove that Lemma 4.1 is true. 5

From Lemma 4.1, we conclude that as long as all the local tracking errors ei (t)
for i = 1, . . . , N go to zero as t → ∞, the global tracking will be achieved.
Substituting (4.9) into the derivative of the local tracking error (4.10), we obtain,

1 X
ėi (t) = ẋi (t) − ẋj (t)
ni v ∈N
j i
   
1 X 1 X
= Ae (xi (t) − xj (t)) + Aj xj (t) + Bj uj (t) − Aj xj (t) + Bj uj (t)
ni v ∈N ni v ∈N
j i j i
 
1 X
= Ae (xi (t) − xj (t))
ni v ∈N
j i
 
1 X
= Ae xi (t) − xj (t) = Ae ei (t) (4.11)
ni v ∈N
j i

Recall Ae is a stable matrix, so (4.11) indicates that limt→∞ ei (t) = 0 exponentially.

 85

4.3.2 Adaptive Control Scheme

In our problem, the leader parameter matrices A0 , B0 and the followers parameter
matrices Ai and Bi (i = 1, . . . , N ) are unknown, so the ideal control protocol (4.8)
will no longer work. An adaptive version protocol of (4.8) is proposed as follows,

1 X T T
ui (t) = K (t)(xi (t) − xj (t)) + K2ij (t)uj (t) + K3ij (t)xj (t) (4.12)
ni v ∈N 1i
j i

with the following adaptive laws

T 1 T T
K̇1ij (t) = − Si Be P ei (t)(xi (t) − xj (t))T (4.13)
ni
1
K̇2ij (t) = − SiT BeT P ei (t)uTj (t) (4.14)
ni
T 1
K̇3ij (t) = − SiT BeT P ei (t)xTj (t), (vj ∈ Ni , i = 1, 2, . . . , N ) (4.15)
ni

∗ ∗ ∗
where K1ij (t), K2ij (t) and K3ij (t) are the estimates of K1i , K2ij and K3ij respectively,
and Si satisfies Assumption (A7) for i = 1, . . . , N ) and P = P T > 0 satisfying
ATe P + P Ae = −Q < 0, Q = QT > 0. K1ij (0), K2ij (0) and K3ij (0) can be chosen
arbitrarily.
∗
Remark 4.2 For each follower agent vi , the ideal controller parameter K1i is
∗
unique. However, according to (4.12), for each different vj ∈ Ni , the estimate K1i s
are different. In order to distinguish these different estimate values, we denote the
∗
estimate K1i as K1ij (t). 

We now present the main result in solving the leader-following consensus tracking
problem for multi-agent systems as follows.

Theorem 4.1. The distributed adaptive controller (4.12), with the adaptive laws
 86

(4.13)-(4.15), applied to multi-agent systems (4.1) guarantees all closed-loop signals

are bounded and global tracking is achieved: limt→∞ (xi (t) − x0 (t)) = 0.

4.4 Stability Analysis

In oder to analyze the stability of the closed-loop system, we first denote

T T ∗T ∗ T T ∗T
K̃1ij (t) = K1ij (t) − K1i , K̃2ij (t) = K2ij (t) − K2ij , K̃3ij (t) = K3ij (t) − K3ij , (4.16)

Substituting the distributed adaptive protocol (4.12) into the multi-agent subsystem
(4.1) leads to

ẋi (t) = Ai xi (t) + Bi ui (t)

1 X
T T


= Ai xi (t) + Bi K1ij (t)(xi (t) − xj (t)) + K2ij (t)uj (t) + K3ij (t)xj (t)
ni
vj ∈Ni
1 X
= (Ae (xi (t) − xj (t)) + Aj xj (t) + Bj uj (t)) (4.17)
ni
vj ∈Ni
1 X 
∗−1 T ∗−1 ∗−1 T

+ Be K4i K̃1ij (t)(xi (t) − xj (t)) + K4i K̃2ij (t)uj (t) + K4i K̃3ij (t)xj (t) .
ni
vj ∈Ni

Then, tracking error equations evolve to

1 X
ėi (t) = ẋi (t) − ẋj (t)
ni v ∈N
j i

1 X
= Ae (xi (t) − xj (t))
ni v ∈N
j i

1 X ∗−1

T T

+ Be K4i K̃1i (t)(xi (t) − xj (t)) + K̃2ij (t)uj (t) + K̃3ij (t)xj (t)
ni v ∈N
j i

= Ae ei (t) (4.18)
1 X ∗−1

T T

+ Be K4i K̃1i (t)(xi (t) − xj (t)) + K̃2ij (t)uj (t) + K̃3ij (t)xj (t)
ni v ∈N
j i
 87

Choose a positive definite function as a Lyapunov candidate to measure the closed-

loop system errors,
N
X
V = Vi (4.19)
i=1

with Vi = Vi1 + Vi2 where

Vi1 = eTi P ei (4.20)

and

X X X
V2i = tr[K̃1i Ms−1 K̃1i
T
]+ T
tr[K̃2ij Ms−1 K̃2ij ] + tr[K̃3ij Ms−1 K̃3ij
T
] (4.21)
vj ∈Ni vj ∈Ni vj ∈Ni

where P = P T > 0, such that ATe P + P Ae + = −Q < 0, Q = QT > 0, and Ms is

∗
a positive symmetric matrix which satisfies Ms = K4i Si , Si ∈ Rp×p (see Assumption
(A4.5)).
Substituting (4.18) into the derivative of (4.20), we have

V̇1 = 2eTi (t)P ėi (t)

= eTi (t)(P Ae + ATe P )ei (t)

2 ∗−1
X
+ eTi (t)P Be K4i T
K̃1i (t)(xi (t) − xj (t))
ni v ∈N j i

2 ∗−1
X
+ eTi (t)P Be K4i K̃2ij (t)uj (t)
ni v ∈N j i

2 T ∗−1
X
T
+ e (t)P Be K4i K̃3ij (t)xj (t) (4.22)
ni i v j ∈Ni

Also, the derivate of V2i becomes to

tr[K̃1i Ms−1 K̃˙ 1i Ms−1 K̃˙ 2ij ]+2 tr[K̃3ij Ms−1 K̃˙ 3ij
X X X
T T T
V̇2i = 2 ]+2 tr[K̃2ij ] (4.23)
vj ∈Ni vj ∈Ni vj ∈Ni
 88

tr[K̃1i Ms−1 K̃˙ 1i

T
P
Substituting (4.13) into vj ∈Ni ], it is obtained

tr[K̃1i Ms−1 K̃˙ 1i

X
T
]
vj ∈Ni

1 X
=− tr[K̃1i Ms−1 SiT BeT P ei (t)(xi (t) − xj (t))]
ni v ∈N
j i

1 X
=− tr[K̃1i K4∗−1 BeT P ei (t)(xi (t) − xj (t))]
ni v ∈N
j i

1 X
=− tr[eTi (t)P Be K4∗−1 K̃1i
T
(xi (t) − xj (t))]
ni v ∈N
j i

1 T X
=− ei (t)P Be K4∗−1 T
K̃1i (xi (t) − xj (t)) (4.24)
ni v ∈N j i

Similarly, it is obtained

1
Ms−1 K̃˙ 2ij ] = − eTi (t)P Be K4i
X X
T ∗−1
tr[K̃2ij K̃2ij (t)uj (t) (4.25)
vj ∈Ni
ni v ∈N j i

1
tr[K̃3ij Ms−1 K̃˙ 3ij
X X
T ∗−1
] = − eTi (t)P Be K4i T
K̃3ij (t)xj (t) (4.26)
vj ∈Ni
ni v ∈N j i

Substituting (4.24)-(3.33) into (4.23) leads to

2 T ∗−1
X
T
V̇2i = − ei (t)P Be K4i K̃1i (t)(xi (t) − xj (t))
ni v ∈N j i

2 T ∗−1
X
− e (t)P Be K4i K̃2ij (t)uj (t)
ni i v j ∈Ni

2 T ∗−1
X
T
− ei (t)P Be K4i K̃3ij (t)xj (t) (4.27)
ni v ∈N j i

From (4.22) and (4.27), we have

V̇i = V̇1i + V̇2i = eTi (t)(P Ae + ATe P )ei (t) = −eTi (t)Qei (t) ≤ 0 (4.28)
 89

Finally, the derivative of the positive definite function V is

N
X N
X N
X
V̇ = V̇i = eTi (t)Qei (t) ≤ −qm kei (t)k22 ≤ 0 (4.29)
i=1 i=1 i=1

where qm > 0 is the minimum eigenvalue of Q.

From (4.29), the desired properties of the proposed adaptive laws are obvious:

• V > 0 and V̇ ≤ 0 implies that the equilibrium state eic = 0 of the closed-loop
system consisting of (4.13) and (4.15) for i = 1, . . . , N is uniformly stable and
its solution eic (t) is uniformly bounded, which gives the boundedness of all xi (t),
K1ij (t), K2ij and K3ij in the multi-agent system, and in turn the boundedness
of ėi (t) for i = 1, . . . , N because of (4.18);

• (4.29) implies ei (t) ∈ L2 for i = 1, . . . , N ;

• with ei (t) ∈ L2 L∞ and ėi (t) ∈ L∞ , applying Barbalat lemma, we con-

T

clude that limt→∞ ei (t) = 0 for i = 1, . . . , N . Then according to Lemma 1,

limt→∞ (xi (t) − x0 (t)) = 0 for i = 1, . . . , N which means the trajectories of all
N follower agents have the desired properties;

4.5 Disturbance Rejection

A distributed adaptive disturbance compensator develops from the basic case in

Chapter 3 is presented in this section. It is proofed by a stability analysis that the
compensator has the capability to make the followers track the leader on the state
and reject the effect of the disturbances as well.
Based on the experience, we have already known that for disturbance rejection,
we develop a new term in the adaptive controller comparing with the disturbance-free
 90

case and chose an appropriate expression for that new term according to which part
does the disturbance act on.
However, in this section, there are N + 1 followers, for each follower i it could be
followed by other agents as well. In order to develop the problem more conveniently,
we re-state the followers structure.
Now consider the dynamic system of all the followers becomes as

xi (t) = Ai xi (t) + Bi ui (t) + Bi di (t) (4.30)

where
qi
X
di (t) = di0 + diβ fiβ (t) (4.31)
β=1

with di0 , diβ ∈ Rpi being some unknown constants and fiβ(t) being some known
bounded continuous functions, β = 1, 2, . . . , qi and some qi > 0. The given leader is
(4.2).
To cancel the disturbance, the nominal control protocol (4.8) is modified as,

 
1 X
u∗i (t) = ∗T ∗ ∗T ∗
K1i (t)(xi (t) − xj (t)) + K2ij (t)uj (t) + K3ij (t)xj (t) + k5ij (t) , (4.32)
ni v ∈N
j i

∗
where the compensation signal k5ij (t) is

∗ ∗
k5ij = K2ij dj (t) − di (t)
qj qi
X X
∗ ∗
= (K2ij dj0 − di0 ) + K2ij djβ fjβ (t) − diβ fiβ (t)
β=1 β=1
qj qi
X X
∗ ∗ ∗
= k50ij + k5jβ fjβ (t) + k5iβ fiβ (t) (4.33)
β=1 β=1
 91

for agent vj ∈ Ni where

∗ ∗
k50ij = K2ij dj0 − di0
∗ ∗
k5jβ = K2ij djβ
∗
k5iβ = −diβ (4.34)

The ideal disturbance compensator (4.32) leads the subsystem (4.1) to a closed-
loop result

1 X ∗T ∗T
ẋi (t) = ((Ai + Bi K1i )(xi (t) − xj (t)) + (Ai + Bi K3ij )xm (t)
ni v ∈N
j i
qj qi
X X
∗ ∗
+ BK2ij uj (t) + Bi K2ij dj0 + Bi K2∗ djβ fjβ (t) − Bi diβ fiβ (t)
β=1 β=1

+ Bi di (t) − Bi di0 )
1 X
= (Ae (xi (t) − xj (t)) + Aj xj (t) + Bj uj (t) + Bj dj (t)) (4.35)
ni v ∈N
j i

In result, we obtain ėi = Ae ei (t). Since Ae is stable, it is easily to verify that

limt→∞ ei (t) = 0 exponentially.
Since the parametric uncertainties, (4.32) does not work for this leader-following
consensus problem, update (4.32) to an adaptive version,

 
1 T T
ui (t) = K1i (t)(xi (t) − xj (t)) + K2ij (t)uj (t) + K3ij (t)xj (t) + k5ij (t) (4.36)
ni

where
qj qi
X X
k5ij (t) = k50ij (t) + k5jβ (t)fjβ (t) + k5iβ (t)fiβ (t) (4.37)
β=1 β=1

∗ ∗
with k50ij (t) being the estimate of k50ij and k5jβ (t) being the estimate of k5jβ . k5iβ (t)
∗ ∗
is the estimate of the k5iβ . k5ij (t) is the estimate of k5ij (t).
 92

The adaptive law for K1i (t), K2ij (t) and K3ij (t) are still given in (4.13), (4.14)
and (4.15). The estimate k5ij is developed as

1 T T
k̇50ij (t) = − S B P ei (t), (4.38)
ni i e
1
k̇5jβ (t) = − SiT BeT P ei (t)fjβ
T
(t), β = 1, . . . , qj , (4.39)
ni
1
k̇5iβ (t) = − SiT BeT P ei (t)fiβ
T
(t), β = 1, . . . , qi . (4.40)
ni

Theorem 4.2. The distributed adaptive controller (4.32), with the adaptive laws
(4.13)-(4.15) and (4.38) - (4.40), applied to multi-agent systems (4.1) guarantees that
all closed-loop signals are bounded and global tracking are achieved: limt→∞ (xi (t) −
x0 (t)) = 0.

Proof: the stability of the closed-loop system, first denote the parameter errors for
k5ij (t),

∗
k̃50ij (t) = k50ij − k50ij (t)
∗
k̃5jβ (t) = k5jβ − k5jβ (t)
∗
k̃5iβ (t) = k5iβ − k5iβ (t) (4.41)

Combining with (4.36), we obtain


1 X ∗−1 T
ėi (t) = Ae ei (t) + Be K4i K̃1i (t)(xi (t) − xj (t)) + K̃2ij (t)uj (t)
ni v ∈N
j i
qj qi 
X X
T
+ K̃3ij (t)xj (t) + k̃50ij (t) + k̃5jβ (t)fjβ (t) + k̃5iβ (t)fiβ (t) (4.42)
β=1 β=1

Choose a positive definite function as a measurement of closed-loop signals of the

 93

multi-agent system
N
X
V̄ = V̄i (4.43)
i=1

with
V̄i = Vi + Vai (4.44)

where

X X X
Vai = T
tr[k̃50ij Ms−1 k̃50ij ] + T
tr[k̃5jβ Ms−1 k̃5jβ ] + T
tr[k̃5iβ Ms−1 k̃5iβ ] (4.45)
vj ∈Ni vj ∈Ni vj ∈Ni

and Vi is in (4.19).
Being similar with (4.24), we obtain

1 T
Ms−1 k̃˙ 50ij ] = −
X X
T ∗−1
tr[k̃50ij ei (t)P Be K4i k̃50ij (t) (4.46)
vj ∈Ni
ni v ∈N j i

1
tr[k̃5jβ Ms−1 k̃˙ 5jβ
X X
T ∗−1
] = − eTi (t)P Be K4i T
K̃5jβ (t)fjβ (t) (4.47)
vj ∈Ni
ni v j ∈Ni

1 T
tr[k̃5iβ Ms−1 k̃˙ 5iβ
X X
T ∗−1 T
]=− ei (t)P Be K4i k̃5iβ (t)fiβ (t) (4.48)
vj ∈Ni
ni v ∈N j i

which helps use to derive the derivative of V as

N
X
V̇ = − eTi (t)Qei (t), Q = QT > 0 (4.49)
i=1

(the details are similar to the proof of Theorem 1 and are omitted here). From this
expression of V̇ , we can similarly obtain the results of Theorem 2. 5
 94

4.6 Simulation Study

Case I: One leader and three followers without disturbance We provide a

simulation example to illustrate the leader-following consensus performance of multi-
agent systems under the proposed adaptive protocol. In particular, a network con-
sisting of three follower agents and one leader is considered. In systems (4.1) and
(4.2),

       
−2 3  0 1  0 2 −1 2
A1 =  , B
 1  = , A
 2  = , B
 2  = ,
1 −2 1 2 −2 1 0 1

      
−1 0  1 0  0 1 1 −2
A3 =   , B3 =   , A0 =   , B0 =  ,
−1 −2 3 1 −1 −3 4 1
  
1 1 1 −2
Ae =   , Be =  , (4.50)
−9 −4 4 1

u0 = [sin(t), cos(t)]T . It can be easily verified that Assumption 4.1 is satisfied. Th

interaction graph of the three agents and the leader is shown in 4.9, which satisfies
the topology precondition Assumption 4.1 and 4.2.
The initial values of the updating matrices and vectors in adaptive laws (4.12)-
(4.15) with K1ij (0), K2ij (0), K3ij (0) are 0.8 times of their ideal values, respectively.
The performance of the leader-following consensus is presented in Figure 4.10. Track-
ing errors xi (t)−x0 (t), i = 1, 2, 3 are shown in Figure 4.11. Obviously, the three agents
can track the leader with reference signal u0 (t).
 95

1 2

3
Figure 4.9: Interaction graph of three follower agents and one leader

 
1 0 0
 
The corresponding matrix Q of Fig 4.1 is  0 1 0.
 
 
 
−1 −1 2

Case II: One leader and five followers without disturbance Consider a multi-
agent systems consisting of five followers and one leader. This case is verified that a
follower can achieve the consensus performance no matter the path length from its
neighbors to the leader are the same or not.
The interaction relationships between the followers and the leader is shown in
Figure 4.4. Leader-following consensus performance is shown in Figure 4.5. Tracking
errors are shown in Figure 4.6. In this case, agents 0-3 have the same dynamic models
with agents 0-3 in Case I. Also Ae and Be are the same. And

       
1 3 −1 2 −1 −4 1 0 
A4 =   , B4 =   , A5 =   , B5 =  , (4.51)
1 1 0 1 3 −2 3 1

.
 96

6
x
1,1
x2,1
4
x3,1
x4,1
2

−2

−4
0 5 10 15 20 25 30 35 40 45 50

10
x1,2
8 x2,2

6 x
3,2
x
4,2
4

−2

−4
0 5 10 15 20 25 30 35 40 45 50

Figure 4.10: State trajectories of three follower states xi and leader state x0 vs.
time(sec)

1
x −x
1,1 0,1
x2,1 −x0,1
0
x3,1 − x0,1

−1

−2

−3

−4
0 5 10 15 20 25 30 35 40 45 50

8
x −x
1,2 0,2
6 x2,2 − x0,2
x3,2 − x0,2
4

−2

−4
0 5 10 15 20 25 30 35 40 45 50

Figure 4.11: Tracking errors between three follower states xi and leader state x0 vs.
time(sec)
 97

2 3 4

0 1 5
Figure 4.12: Interaction graph of five follower agents and one leader

 
1 0 0 0 0
 
0 1 0 0 0
 
 
 
The corresponding Q of Fig 4.4 = 
−1 −1 2 0 0

 
 
−1 0 −1 2 0
 
 
−1 0 0 0 1

Case III: One leader and three followers with disturbance Multi-agent sys-
tem in this case is a system which has disturbances acting on multi-agent system in
Case I in Section 4.6. Parameters matrices of each agents including all the followers
and the leader are shown in (4.50). Disturbances acting on agent 1-agent 3 are

d1 = [−5, −5]T + [1, 2]T sin(t)

d2 = [−4, −4]T + [2, 1]T cos(2t)

d3 = [2, 6]T + [6, 6]T sin(3t) (4.52)

respectively. u0 = [sin(t), cos(t)]T . K1ij (0), K2ij (0), K3ij (0) and k5ij (0) are 0.85 times
their ideal value.
Use adaptive disturbance compensator (4.4) we can get the desired state trajecto-
ries. Leader-following consensus performance is shown in Figure 4.7. Tracking errors
are shown in Figure 4.8.
 98

6 x
0,1
4 x1,1
x2,1
2
x3,1
0 x4,1
x5,1
−2

−4

−6
0 10 20 30 40 50 60 70 80

5
x0,2
x1,2
x2,2
x3,2
0 x
4,2
x5,2

−5
0 10 20 30 40 50 60 70 80

Figure 4.13: State trajectories of five follower states xi and leader state x0 vs.
time(sec)

4
x1,1 − x0,1
x2,1 − x0,1
2 x −x
3,1 0,1
x4,1 − x0,1
0 x5,1 − x0,1

−2

−4
0 10 20 30 40 50 60 70 80

3
x1,2 − x0,2
2 x2,2 −x0,2
x3,2 − x0,2
1
x4,2 −x0,2
0 x5,2 − x0,2

−1

−2

−3
0 10 20 30 40 50 60 70 80

Figure 4.14: Tracking errors between five follower states xi and leader state x0 vs.
time(sec)
 99

8
x0,1
6
x1,1
4 x2,1

2 x3,1

−2

−4

−6
0 5 10 15 20 25 30 35 40 45 50

8
x0,2
6 x1,2
x2,2
4
x3,2
2

−2

−4
0 5 10 15 20 25 30 35 40 45 50

Figure 4.15: State trajectories of three follower states xi and leader state x0 with
disturbance rejection vs. time(sec)
.
1

−1

x1,1 − x0,1
−2 x2,1 − x0,1
x3,1 − x0,1
−3
0 5 10 15 20 25 30 35 40 45 50

8
x1,2−x0,2
6 x2,2 − x0,2
x3,2 − x0,2
4

−2

−4
0 5 10 15 20 25 30 35 40 45 50

Figure 4.16: Tracking error between follower states xi and leader state x0 vs. time(sec)
Chapter 5

Conclusions and Future Work

5.1 Conclusions

In this thesis, in order to solve the distributed leader-following problems for the
multi-agent systems, we first solved the one leader-one follower problem with and
without the disturbances in Chapter 3. Stability analyses have shown that the pro-
posed controllers in Chapter 3 can achieve the desired properties and the simulation
studies also verify the capability of the proposed adaptive control scheme, i.e., all
agents can track the prescribed leader eventually. Based on the basic idea used in
Chapter 3, we have developed two solutions to the distributed adaptive control and
disturbance compensation problems for multi-input multi-agent systems in Chapter 4.
Our study has shown that the desired closed-loop system stability and tracking prop-
erties can be achieved by control adaptation based on a complete parameterization of
the leader and the followers system parameters as well as the disturbance parameters.
Simulation results for multi-agent systems have also confirmed the capability of the
proposed adaptive control.

100
 101

5.2 Future Research Topics

Adaptive control for distributed multi-agent coordination with leader and followers
parametric uncertainties, as discussed and analyzed in this thesis, is a new research
focus. However, there are still many challenges need to be overcame. In this section,
we will introduce several potential extensions to our research.
The control framework discussed in Chapter 3 assumes that all of the states of
each agent are available for feedback control design and exchange in a communication
network. However, in real application, not all of the states can be accessible for
measurement and exchange through a communication network. It is both practically
and theoretically important to relax the requirement for full state measurement or
exchange for multi-agent systems. Therefore,a state observer is needed to achieve
the state tracking. Also this useful extension results can be applied to the multi-
agent case in Chapter 4. It is worthwhile investigating the estimation of relative state
estimation with appropriate estimation accuracy as well as the stability of closed-loop
multi-agent systems with estimated information.
At the meanwhile, for some practical multi-agent systems, links between different
agents have different weights. It is worth and important to figure out how to make
the multi-agent consensus well based on those different weights. Compared to the
weighted case, in our thesis, we assume that every edge in the communicate graph
have the same weight. This challenge need to be solved for many real practical
applications.
Bibliography

[1] Bernstein, D. S., “A student’s guide to classical control,” IEEE Control Systems
Magazine, vol. 17, no. 4, pp. 96-100, 1997.

[2] Chen, Chi-Tsong, Linear System Theory and Design, Oxford University Press
Inc., 1995.

[3] Davison, E., “The robust control of a servomechanism problem for linear time-
invariant multivariable systems,” IEEE Transactions on Automatic Control, vol.
21, no. 1, pp. 25-34, 1976.

[4] Fax, J. A. and R. M. Murray, “Information flow and cooperative control of vehicle
formations,” IEEE Transactions on Automatic Control, vol. 49, no. 9, pp. 1465-
1476, 2004.

[5] Francis, B. A. and W. M. Wonham, “The internal model principal of control

theory,” Automatica, vol. 12, no. 5, pp. 457-465, 1976.

[6] Godsil, C. and G. Royle, Algebraic Graph Theory (Vol. 207), Springer, New York,
2001.

[7] Gross, J. L. and J. Yellen, Handbook of Graph Theory, CRC Press, Boca Raton,
FL, 2004.

102
 103

[8] Lafferriere, G., A. Williams, J. Caughman and J. J. P. Veerman, “Decentralized

control of vehicle formations,” Systems and Control Letters, vol. 54, no. 9, pp.
899-910, 2005.

[9] Hong, Y., J. Hu and L. Gao, “Tracking control for multi-agent consensus with
an active leader and variable topology,” Automatica, vol. 42, no. 7, pp. 1177-1182,
2006.

[10] Hong, Y., G. Chen and L. Bushnell, “Distributed observers design for leader-
following control of multi-agent networks,” Automatica, vol. 44, no. 3, pp. 846-850,
2008.

[11] Hong, Y., X. Wang and Z. Jiang, “Distributed output regulation of leaderfollower
multi-agent systems,” International Journal of Robust Nonlinear Control, vol. 23,
no. 1, pp. 48-66, 2013.

[12] Hu, J. and G. Feng, “Distributed tracking control of leader-follower multi-agent

systems under noisy measurement,” Automatica, vol. 46, no. 8, pp. 1382-1387, 2010.

[13] Jadbabaie, A., J. Lin and A. S. Morse, “Coordination of groups of mobile au-
tonomous agents using nearest neighbor rules,” IEEE Transactions on Automatic
Control, vol. 46, no. 6, pp. 988-1001, 2003.

[14] Ji, M. and M. Egerstedt, “A graph-theoretic characterization of controllability

for multi-agent systems,” American Control Conference, 2007. Proceedings of the
2007, pp. 4588-4593, 2007.

[15] Khalil, H. K., Nonlinear Systems, Prentice Hall Inc., 3rd Ed., Upper Saddle
River, NJ, 2002.
 104

[16] Lavretsky, E. and K. A. Wise, Robust and Adaptive Control with Aerospace Ap-
plication, Springer, London, United Kingdom, 2013.

[17] Li, T. and J. Zhang, “Mean square average consensus under measurement noises
and fixed topologies: necessary and sufficient conditions,” Automatica, vol. 45, no.
8, pp. 1929-1936, 2009.

[18] Li, Z., Z. Duan and L. Huang, “H∞ control of networked multi-agent systems,”
Journal of Systems Science and Complexity, vol. 22, no. 1, pp. 35-48, 2009.

[19] Lin, P., Y. Jia, “Robust H∞ consensus analysis of a class of second-order multi-
agent systems with uncertainty,” IET Control Theory and Applications, vol. 4, no.
3, pp. 487-498, 2010.

[20] Li, Z., Z. Duan and G. Chen, “Dynamic consensus of linear multi-agent systems”,
IET Control Theory and Applications, vol. 5, no. 1, pp. 19-28, 2011.

[21] Li, Z., X. Liua, W. Ren and L. Xie, “Distributed tracking control for linear
multi-agent systems with a leader of bounded unknown input,” IEEE Transactions
on Automatic Control, vol. 58, no. 2, pp. 518523, 2013.

[22] Liu, B., T. Chu, L. Wang and G. Xie, “Controllability of a leader-follower dy-
namic network with switching topology,” IEEE Transactions on Automatic Control,
vol. 53, no. 4, pp. 1009-1013, 2008.

[23] Liu, Y. and Y. Jia, “H∞ consensus control of multi-agent systems with switching
topology: a dynamic output feedback protocol,” International Journal of Control,
vol. 83, no. 3, pp. 527537, 2010.
 105

[24] Liu, Y. and Y. Jia, “Adaptive leader-following consensus control of multi-agent

systems using model reference adaptive control approach,” IET Control Theory
and Applications, vol. 6, no. 13, pp. 2002-2008, 2012.

[25] Luo, J., Adaptive Control for Distributed Multi-Agent Coordination, Ph.D. The-
sis, University of Connecticut, 2013.

[26] Meng, Z., W. Ren, Y. Cao and Z. You, “Leaderless and leader-Following con-
sensus with communication and input delays under a directed network topology,”
IEEE Transactions on Systems, Man and Cybernetics-Part B: Cybernetics, vol. 41,
no, 1, pp. 75-88, 2011.

[27] Moreau, L., “Stability of multi-agent systems with time-dependent communica-

tion links,” IEEE Transactions on Automatic Control, vol. 50, no. 2, pp. 169-182,
2005.

[28] Ni, W. and D. Cheng, “Leader-following consensus of multi-agent systems under

fixed and switching topologies,” Systems and Control Letters, vol. 59, no. 3, pp.
209-217, 2010.

[29] Nise, N. S., Control Systems Engineering, John Wiley and Sons Inc., 6th Ed.,
Hoboken, NJ, 2011.

[30] Peng, Z., D. Wang, H. Zhang, G. Sun and H. Wang, “Distributed model refer-
ence adaptive control for cooperative tracking of uncertain dynamical multi-agent
systems, IET Control Theory and Applications, vol. 7, no. 8, pp. 1079-1087, 2013.

[31] Ren, W. and R. W. Beard, “Consensus seeking in multi-agent systems under dy-
namically changing interaction topologies, IEEE Transactions on Automatic Con-
trol, vol. 50, no. 5, pp. 655-661, 2005.
 106

[32] Ren, W., R. W. Beard and E. M. Atkins, “A survey of consensus problems in

multi-agent coordination,” Proceedings of the 2005 American Control Conference,
vol. 3, pp. 1859-1864, Portland, OR, 2005.

[33] Ren W., “On consensus algorithms for double-integrator dynamics,” IEEE
Transactions on Automatic Control, vol. 53, no. 6, pp. 1503-1509, 2008.

[34] Olfati-Saber, R. and R. M. Murray, “Distributed cooperative control of multi-

agent vehicle formations using structural potential functions,” Proceedings of the
15th IFAC World Congress, pp. 346-352, Barcclona, Spain, 2002.

[35] Olfati-Saber, R. and R. M. Murray, “Consensus problems in networks of agents

with switching topology and time-delays”, IEEE Transactions on Automatic Con-
trol, vol. 49, no. 9, pp. 1520-1533, 2004.

[36] Olfati-Saber, R., “Flocking for multi-agent dynamic systems: Algorithms and
theory.” IEEE Transactions on Automatic Control, vol. 51, no. 3, pp. 401-420,
2006.

[37] Olfati-Saber, R., J. A. Fax and R. M. Murray, “Consensus and cooperation in

networked multi-agent systems,” Proceedings of the IEEE, vol. 95, no. 1, pp. 215-
233, 2007.

[38] Slotine, J. J. E. and W. Li, Applied Nonlinear Control, Prentice Hall Inc., Upper
Saddle River, NJ, 1991.

[39] Su, H., X. Wang and Z. Lin. “Flocking of multi-agents with a virtual leader,”
IEEE Transactions on Automatic Control, vol. 54, no. 2, pp. 293-307, 2009.
 107

[40] Su, Y. and J. Huang, “Cooperative output regulation of a linear multi-agent

system,” IEEE Transactions on Automatic Control, vol. 57, no. 4, pp, 1062-1066,
2012.

[41] Tanner, H. G., A. Jadbabaie and G. J. Pappas, “Flocking in fixed and switching
networks,” IEEE Transactions on Automatic Control, vol. 52, no. 5, pp. 863-868,
2007.

[42] Tao, G., Adaptive Control Design and Analysis, John Wiley and Sons Inc., New
York, NY, 2003.

[43] Tang, Y., “Leader-following coordination problem with an uncertain leader in

a multi-agent system,” IET Control Theory and Applications, vol. 8, no. 10, pp.
773-781, 2014.

[44] Xiao, L., S. Boyd and S. Kim, “Distributed average consensus with least-mean-
square deviation,” Journal of Parallel and Distributed Computing, vol. 67, no. 1,
pp. 33-46, 2007.

[45] Xiao, F. and L. Wang, “Asynchronous consensus in continuous-time multi-agent

systems with switching topology and time-varying delays,” IEEE Transactions on
Automatic Control, vol. 53, no. 8, pp. 1804-1816, 2008.

[46] Wang, J., D. Cheng and X. Hu, “Consensus of multi-agent linear dynamic sys-
tems,” Asian Journal of Control, vol. 10, no. 2, pp.144-155, 2008.

[47] Weiss G., Multiagent systems: Intelligent robotics and autonomous agents, MIT
Press, Cambridge, MA, 2013.

[48] Wonham, W. M., Linear Multivariable Control: A Geometric Approach,

Springer-Verlag, New York, NY, 1979.
 108

[49] Wooldridge, M. and N. R. Jennings, “Intelligent agents: theory and practice,”

The Knowledge Engineering Review, vol. 10, no. 2, pp.115-152, 1995.

[50] Zhang, H. and F. L. Lewis “Adaptive cooperative tracking control of higher-

order nonlinear systems with unknown dynamics,” Automatica, vol. 48, no. 7, pp.
1432-1439, 2012.

[51] Zhu, W. and D. Cheng, “Leader-following consensus of second-order agents with

multiple time-varying delays,” Automatica, vol. 46, no. 12, pp. 1994-1999, 2010.