Professional Documents
Culture Documents
Leader-Following Consensus Protocol For Second-Order Multi-Agent Systems Using Neural Networks
Leader-Following Consensus Protocol For Second-Order Multi-Agent Systems Using Neural Networks
Abstract: In this paper, an adaptive leader-following consensus protocol is proposed for second-order multi-agent systems with
unknown nonlinear dynamics based on neural networks (NNs), and it is proved that if at least one agent in each connected
component of the fixed interaction graph is connected to the leader, all agents can asymptotically track the leader. By estimating
the tracking error in terms of L2 norm, the transient performance of multi-agent systems can be significantly improved. Finally,
a numerical example is included to the proposed protocol.
Key Words: Multi-agent systems, Nonlinear dynamics, Leader-following consensus protocol, Neural networks
535
2.2 Algebraic Graph Theory Let x̄i (t) = xi − x0 and v̄i (t) = vi − v0 , then we can obtain the
Undirected graphs are applied to model the interaction error dynamics of (1) as follows:
topologies among agents. Let G = (V , E , A) be a weighted
x̄˙i = v̄i
undirected graph of order n with the set of nodes V = (4)
v̄˙i = a(ξi ) + bui − a0
{w1 , w2 , · · · , wn }, the set of unordered pairs of nodes (called
edges) E ⊆ V × V , and a weighted adjacency matrix A =
Under the assumption that the nonlinear term a(ξi ) and the
[ai j ] with nonnegative adjacency elements ai j = a ji . In graph
constant b are known, one can state the following result.
G , node wi represents the ith agent, and ei j = (wi , w j ) (or e ji )
denotes an undirected edge between agents i and j. Then
Lemma 2 Suppose the interaction graph G is connected.
the set of neighbors of wi is denoted by Ni = {w j ∈ V :
For the ith system (4), if the protocol is designed as
(w j , wi ) ∈ E }. Graph G is called connected if there is an
undirected path between any pair of distinct nodes wi and
w j , ∀(i, j) ∈ N {1, 2, · · · , n}.
ui = ∑ {k1[ai j (x j − xi ) + ci(x0 − xi)]
j∈Ni
(5)
Given the weighted adjacency matrix A = [ai j ] with ai j 0, a(ξi ) − a0
+k2 [ai j (v j − vi ) + ci (v0 − vi )]} −
it is stipulated that e ji ∈ E (or ei j ∈ E ) ⇔ ai j = a ji > 0, b
i.e., ai j = a ji > 0 if and only if information flows between
agents i and j. A diagonal matrix D = diag(d1 , d2 , · · · , dn ) is where consensus gains k1 and k2 satisfy k2 > 1/(bλ ) + k1 /4,
named the degree matrix of G , whose diagonal elements are then xi (t) → x0 as t → ∞ (∀i ∈ N ).
di = ∑nj=1 ai j . Then the graph Laplacian of a weighted graph
Proof Applying protocol (5) to the ith system (4) will result
G is defined as L = D − A. Obviously, L1 = 0 and 1T L = 0T
in the closed-loop system
hold.
In what follows, we mainly concern the graph G who is as- x̄˙i = v̄i
sociated with the system consisting of n agents (related to v̄˙i = b ∑ {k1 [ai j (x j − xi ) + ci (x0 − xi )]
graph G ) and one leader. By “graph G of this system is con- j∈Ni
nected”, we mean that at least one agent in each connected +k2 [ai j (v j − vi ) + ci (v0 − vi )]} (6)
component of G is connected to the leader. Denote ci 0
as the connection weight between agent i and the leader, and =b ∑ {k1 [ai j (x̄ j − x̄i) − cix̄i ]
j∈Ni
it is stipulated that ci > 0 if and only if agent i is connected +k2 [ai j (v̄ j − v̄i ) − ci v̄i ]}
to the leader. Let H = L + C be the symmetric matrix as-
sociated with G , where L is the graph Laplacian of G and that can be written in the following matrix form
C = diag(c1 , c2 , · · · , cn ).
0n In
η̇ = η
Lemma 1 [10] If graph G is connected, then the symmetric −bk1 H −bk2 H (7)
matrix H associated with G is positive definite. Fη
If matrix H is positive definite, then denote λ̄ > 0 and λ > 0 where η = [x̄T , v̄T ]T with x̄ = [x̄1 , x̄2 , · · · , x̄n ]T and v̄ =
as the largest and smallest eigenvalues of H, respectively. [v̄1 , v̄2 , · · · , v̄n ]T , and H = L + C has been defined in 2.2.
Consider the Lyapunov function
2.3 Approximation Property of Neural Networks
Let f (χ ) : R p → Rq be a smooth function, then it was V = η T Pη
demonstrated that, as long as χ is restricted to a compact set
S ∈ R p , for some sufficiently large number of hidden-layer with positive definite matrix
neurons, there exist weights and thresholds such that
bk2 H In 1
P= (bk2 > ) (8)
f (χ ) = W T Φ(χ ) + ε (χ ) (3) In In λ
where χ is the input vector, Φ(•) ∈ RL is the NN basis func- Taking the time derivative of V yields
tion, W is the weight matrix of the output layer and ε (χ )
is the approximation error. The integer L denotes the num- V̇ = η T (PF + F T P)η
(9)
ber of the hidden layer neurons. For any choice of a pos- η T Qη
itive number εN , a neural network can be found such that
ε (χ ) εN for all χ ∈ S. where
−2bk1H −bk1 H
Q=
3 MAIN RESULTS −bk1 H 2(In − bk2H)
In this section, a leader-following consensus protocol is first Since graph G is connected, matrix H is positive definite
designed for system (1) in the case that a(ξi ) and b are both due to Lemma 1. Then by the Schur complement theo-
known. Subsequently, based on neural networks, a robustify- rem, Q is negative definite if k2 > 1/(bλ ) + k1 /4, which
ing term is added to construct a protocol, such that all agents can also guarantee that matrix P is positive definite from (8).
asymptotically follow the leader if the interaction graph G is Therefore, it is derived that limt→∞ η (t) = 0, which leads to
connected. limt→∞ xi (t) = x0 for ∀i ∈ N .
536
Consensus protocol (5) can also be expressed as Theorem 1 Consider the ith system (4) with adaptive pro-
tocol (12). If the interaction graph G is connected, and
ui = ∑ {k1[ai j (x j − xi) + ci(x0 − xi)] the positive consensus gains k1 and k2 in (12) satisfy k2 >
j∈Ni (10) λ̄ /bλ 2 + k1 /4, then limt→∞ xi (t) = x0 for ∀i ∈ N .
+k2 [ai j (v j − vi ) + ci (v0 − vi )]} − udi
with udi = (a(ξi ) − a0 )/b. In the case that the nonlinear term Proof Substituting protocol (12) into (4) yields
a(ξi ) and the constant b are both unknown, the desired con-
trol term udi is not available. In the following, neural net-
works will be used to approximate the unknown function udi . x̄˙i = v̄i
Since udi is a smooth function on a compact set Ωξi , it can v̄˙i = a(ξi ) − a0 + b · [Δi1 − ŴiT Φi (ξi ) − ε̂Mi · sgn(−Δi2 )]
be approximated by NNs provided in (3) = a(ξi ) − a0 + b · [udi − udi + Δi1 − ŴiT Φi (ξi )
a(ξi ) − a0 −ε̂Mi · sgn(−Δi2 )]
udi =
b (11) = b · [udi + Δi1 − ŴiT Φi (ξi ) − ε̂Mi · sgn(−Δi2 )]
= WiT Φi (ξi ) + εi (ξi ) = b · [Δi1 + W̃iT Φi (ξi ) + εi (ξi ) − ε̂Mi · sgn(−Δi2 )]
(14)
where Φi (ξi ) = [φi1 (ξi ), φi2 (ξi ), · · · , φiLi (ξi )]T is the basis
which can be written in the matrix form:
vector with Li denoting the NN node number of the ith sys-
tem, Wi ∈ RLi is the weight vector, and εi (ξi ) is the approxi-
0n In 0
mation error. η̇ = η+
−bk1H −bk2 H W +Ξ (15)
Assumption 2 On the compact set Ωξi , there exist an inte- Fη + g
ger Li and an ideal constant wight vector Wi , such that ap-
proximation error satisfies |εi (ξi )| εMi , where εMi > 0 is a where W = [W̃1T Φ1 (ξ1 ), · · · , W̃nT Φn (ξn )]T and Ξ = [ε1 (ξ1 ) −
given constant. ε̂M1 · sgn(−Δ1 2 ), · · · , εn (ξn ) − ε̂Mn · sgn(−Δn 2 )]T are column
vectors.
Generally, the ideal weight vector Wi is unknown and needs
to be estimated. Let Ŵi be the estimate of Wi , then the weight Since graph G is connected, matrix H is positive definite due
estimation error is W̃i = Wi − Ŵi . to Lemma 1. Take the Lyapunov function
Under the Assumptions 1 and 2, design the NN-based adap-
tive consensus protocol as follows: V = V1 + V2
Lemma 3 (Barbalat) If the differentiable function f (t) has According to the Schur complement theorem, T is negative
a finite limit as t → ∞, and if f˙ is uniformly continuous, then definite if k2 > λ̄ /bλ 2 + k1 /4, which also guarantees that
f˙(t) → 0 as t → ∞. matrix R is positive definite due to (16).
537
Note that Theorem 2 For the ith system (4) with adaptive protocol
(12), the L2 bound of the tracking error x̄i is
(x̄T + v̄T)H(W + Ξ)
n T 2
n 1 W̃i0 W̃i0 ε̃Mi0
= ∑ [ ∑ (ai j (x̄i − x̄ j ) + cix̄i ) + (ai j (v̄i − v̄ j ) + ci v̄i )] x̄i 2 μ̄ η0 η0 + ∑
T
+ (22)
i=1 j∈N γ γwi γε i
i
i=1
· W̃iT Φi (ξi ) + εi (ξi ) − ε̂Mi · sgn(−Δi2 )
n where μ̄ > 0 is the largest eigenvalue of matrix R and γ > 0
= ∑ [ ∑ (ai j (xi − x j ) + ci (xi − x0 )) + (ai j (vi − v j ) is the smallest eigenvalue of matrix −T ; η0 , W̃i0 and ε̃Mi0 are
i=1 j∈Ni the initial values of η , W̃i and ε̃Mi , respectively.
+ci (vi − v0 ))] · W̃iT Φi (ξi ) + εi (ξi ) − ε̂Mi · sgn(−Δi2 )
n Proof From (20), it follows that
= ∑ (−Δi2 ) · W̃iT Φi (ξi ) + εi (ξi ) − ε̂Mi · sgn(−Δi2 ) n
i=1
n V̇ η T T η −γ η T η = −γ ∑ (x̄2i + v̄2i ) (23)
∑ −Δi2 · W̃iT Φi (ξi ) + | − Δi2| · εMi − | − Δi2| · ε̂Mi i=1
i=1
n which results in
=∑ −Δi2 · W̃iT Φi (ξi ) + |Δi2 | · ε̃Mi n
i=1 γ x̄2i γ ∑ (x̄2i + v̄2i ) −V̇ (24)
(18) i=1
Then, it is derived that Thus, it yields that
∞
1 ∞
n
x̄i 22 = |x̄i (t)|2 dt − V̇ dt
V̇ η T T η + 2 ∑ (−Δi2 · W̃iT Φi (ξi ) + |Δi2| · ε̃Mi ) 0 γ 0
i=1 1 1 1
W̃ TŴ˙ i ε̃Mi ε̂˙Mi
n = V (0) − V (∞) V (0)
−2 ∑ ( i + ) γ γ γ
γwi γε i n T 2
i=1 1 T W̃i0 W̃i0 ε̃Mi0 (25)
n ˙ (19) = η R η0 + ∑ +
= η T T η + 2 ∑ [W̃iT (−Δi2 Φi (ξi ) −
Ŵ i
) γ 0 i=1 γwi γε i
γ
i=1 wi n
W̃ TW̃ ε̃ 2
1
ε̂˙Mi μ̄ η0T η0 + ∑
i0
i0
+ Mi0
+ε̃Mi (|Δi2 | − )] γ γwi γε i
γε i i=1
538
/HDGHU tems in undirected networks with fixed topology. By apply-
ing NNs to approximate the unknown nonlinear dynamics, a
new adaptive consensus protocol using only local informa-
tion has been designed. Under this NN-based protocol, it
has been proved that all agents can asymptotically track the
Fig. 1 Interaction graph. leader, and several methods have been provided to improve
the transient tracking performance of multi-agent systems.
4.1 Effect of Initial States on the System Performance However, the system model considered in this manuscript is
relatively ideal, and how to design adaptive consensus pro-
Let k1 = 1, k2 = 20, γwi = γε i = 1, and the initial state of the tocols for multi-agent systems with random measurement
leader be ξ0 (0) = [0, 0]T. The initial states of three agents noises may be the future work.
are set ξ1 (0) = [−0.5, 0]T, ξ2 (0) = [0, 0]T , ξ3 (0) = [0.5, 0]T
and ξ1∗ (0) = [−1, 0]T , ξ2∗ (0) = [0, 0]T, ξ3∗ (0) = [1, 0]T re- REFERENCES
spectively, then the corresponding initial error vectors are [1] JADBABAIE A, LIN J, MORSE A S. Coordination of groups
η (0) = [−0.5, 0, 0.5, 0, 0, 0]T and η ∗ (0) = [−1, 0, 1, 0, 0, 0]T. of mobile autonomous agents using nearest neighbor rules.
From Fig. 2, it can be observed that larger transient tracking IEEE Trans. on Automatic Control, 2003, 48(6): 988-1001.
errors exist during the first 10 seconds when the initial states [2] REN W, BEARD R W. Consensus seeking in multiagent
are ξ ∗ than ξ , which demonstrates that small initial value systems under dynamically changing interaction topologies.
η0T η0 leads to small tracking errors during the initial period IEEE Trans. on Automatic Control, 2005, 50(5): 655-661.
of adaptation. [3] OLFATI-SABER R, MURRAY R M. Consensus problems in
networks of agents with switching topology and time-delays.
3
ξ1(0)=[−0.5,0], ξ2(0)=[0,0], ξ3(0)=[0.5,0]
agent 1
3
ξ1(0)=[−1,0], ξ2(0)=[0,0], ξ3(0)=[1,0]
agent 1
IEEE Trans. on Automatic Control, 2004, 49(9): 1520-1533.
2.5
agent 2
agent 3
leader
2.5
agent 2
agent 3
leader
[4] LIU Y, JIA Y, DU J, YU F. Average-consensus problem in
2
2 multiagent systems. Proceedings of the European Control
1.5
Conference, 2007: 889-895.
Position(xi)
Position(xi)
1.5
1
1
[5] LIU Y, JIA Y. Necessary and sufficient conditions for consen-
0.5
0.5
sus in second-order multi-agent systems with fixed topology
0
0
and time-delays. Proceedings of International Symposium on
−0.5
Position(xi)
1.5
1
sign for large-scale systems with strong interconnections us-
0.5
0.5 ing neural networks. IEEE Trans. on Automatic Control,
0
0
2003, 48(5): 805-810.
−0.5
0 5 10 15 20 25 30 35 40
−0.5
0 5 10 15 20 25 30 35 40
[13] LIU W, JAGANNATHAN S, WUNSCH D C, CROW M L.
Time(t) Time(t)
Decentralized neural network control of a class of large-scale
Fig. 3 Performances of three agents when γwi = γε i = 1 and systems with unknown interconnections. IEEE Conference
γwi = γε i = 10, respectively. on Decision and Control, 2004: 4972-4977.
5 CONCLUSION
In this paper, we have considered the leader-following con-
sensus problem for second-order nonlinear multi-agent sys-
539