Complex Dynamical Networks: Lecture 8: Network Control
EE 6605
Instructor: G Ron Chen
Most pictures on this ppt were taken from un-copyrighted websites on the web with thanks
Motivational Examples
Motivation: Example C. elegans
In its Neural Network:
Neurons: ~ 300 Synapses: ~ 3000
Excerpt
The worm Caenorhabditis elegans has 297 nerve cells. The neurons switch one another on or off, and, making 2345 connections among themselves. They form a network that stretches through the nematode’s millimeter-long body.
How many neurons would you have to commandeer in order to control the network with complete precision?
The answer is: 49
— Adrian Cho, Science, 13 May 2011, vol. 332, p 777 Here, control = stimuli
Another Example
“ … very few individuals (approximately 5%) within honeybee swarms can guide the group to a new nest site.”
I.D. Couzin et al., Nature, 3 Feb 2005, vol. 433, p 513
These 5% of bees can be considered as “controlling” or “controlled” agents
Leader-Follower networks
Now …
o Given a network of dynamical systems (e.g., ODEs)
o Given a specific control objective
(e.g., synchronization)
o Assume: a certain class of controllers (e.g., local state- feedback controllers) are chosen to use
Control Problem:
Pining Control:
How many controllers to use?
Where to put them (where to “pin”)?
State Feedback Control
Attempts to Answer
First, consider undirected and unweighted networks
Each node is a higher- dimensional nonlinear dynamical system:
dxi f(x), xRn, i1,2,…,N
o Regular networks
o Random-graph networks o Small-world networks
o Scale-free networks
o …….
dt
i
Network Model
Linearly coupled network:
a general assumption is that f (x) is Lipschitz (e.g., linear: Ax)
dx N
i f(xi)caijHxj xRn
i1,2,…,N
dt j1 i
coupling strength c > 0 and coupling matrices (undirected):
A[a ] ijNN
A: If node
Laplacianmatrix: LDA
r 0 11
r
i connects to node
0r
nn
j (j ≠ i), then aij = aji = 1; else, aij = aji = 0; aii 0 Ddiag{d1,…,dN} di -degreeofnode i
H
22
What kind of controllers? How many? Where?
i1,2,…,N i1,2,…,N
dx
dt j1
N if(xi)c aHx
ij j
+ ui
(u Hx) ii
dx N
i f(xi)caHxHx dt ij j i i
j1
1 if tocontrol
i 0 if notcontrol
Q:Howmanyi 1?Which i?
How Many? – One, or a few, if … c is large enough
dx
1 f(x)c
N
dt 1 if(xi)c aHx
dx
dt j1
N
ij j
i 2,3,…, N
j1
a Hx u u Hx 1jj111
X.F. Wang, G.R. Chen, Physica A (2002): Let l =1 and s = 0 therein
Still yet,
where to apply controllers makes a difference Pinning Control: (How many and which nodes to pin?)
Only a small fraction of nodes are selected for control:
1. Selective pinning scheme 2. Random pinning scheme
X.F. Wang, G.R. Chen, Physica A (2002)
Example
Consider a scale-free CNN, which has a zero equilibrium (s = 0): N
dxi1 j1
xi3xi4c ax
dt 2xi2xi3cN ax
dxi2 j1
ij j1
dxi dt14xi114xi2cN ax
ij j2
dt dxi3 j1
dt 100xi1 100xi4
dxi4 100(xi4 1 xi4 1)
dt
ij j3
cNax ij j4
j1 i 1,2,…, N 60
1 Selective Pinning Control
Here, network size N = 60, coupling strength c = 8 and
number of controlled nodes is
Pin the first 15 largest nodes:
The controlled state x1
m =15, by ui hxi
Control gain: Settling time = 10
h30
2 Random Pinning Control Randomly pin 15 nodes.
Comparison:
The controlled state x1
1. Control gain is much larger: h 513
Recall the last one:
h30
2. It takes twice longer time
to synchronize the network:
Settling time = 20 Recall the last one: 10
Network Controllability Theory
In retrospect, …
(1930-2016)
18
System Controllability
Linear Time-Invariant (LTI) system
x(t) Ax(t) Bu(t)
xRn : statevector uRp : controlinput
ARnn : systemmatrix BRnp : controlmatrix
x1(t) x(t)
x2(t) x(t0)
x3(t)
Controllable: The system orbit can be driven by an input from any initial state to any target state in finite time.
Example
x(t) Ax(t) Bu(t) xax
1 111
x a x a x bu, 2 211 222
x3 a32x2 a33x3
a 0 0 0
11 Aa21 a22 0 , Bb
0aa 0 32 33
Classical Controllability Theorem
x(t) Ax(t) Bu(t) Rank Criterion:
The linear system is controllable if and only if the controllability matrix C has full row rank:
Previous
Example
uncontrollable
What About Directed Networks?
In retrospect: large-scale systems theory Structural Analysis of Dynamical Systems
Q:
Is this kind of structures controllable?
Structural Controllability
Corresponding linearized system has the following general form:
xax 1 111
xaxax bu, 2 211 222
a 0 0 0 11
Aa21 a22 0, Bb 0aa 0
x3 a32x2 a33x3
rankB,AB,A2Bb a b a2 b 2
32 33 00 0
22 22
0 a b a (a a )b 32 32 22 33
Uncontrollable
1 0 0 b0a 0
1 0 0 b0a0
1 0 0 1 0 0 b0a 0b0aaa
1 21 0 0 a a
1 21 0 a 0
1 21 1 21 2331 0 a a a 0 a a a
3221, rank C 3 N
31 , rank C 2 Nn 3
31 3331, rank C 3 Nn
31 3221 rank C ? controllable ?
Examples: Structure matters
C=[B,AB,A2 B]
controllable
uncontrollable
controllable
See also: Y.Y. Liu, J.J. Slotine, A.L. Barabási, Nature (2011)
Building Blocks
Cactus is the minimum structure which contains no inaccessible nodes and no dilations
Structural Controllability Theorem
The following statements are equivalent:
Algebra:
1. An LTI control system (A,B) is controllable Geometry:
2. The digraph G (A,B) contains neither inaccessible nodes nor dilations, but is spanned by a cactus
C.T. Lin, IEEE Trans. Auto. Contr., 1977
Undirected network
Matching:
a set of edges without common nodes
Matched Unmatched node node
Maximum matching
(Perfect matching):
matching of largest size
Matching in Undirected Networks
L. Lovász, M.D. Plummer, Matching Theory (2009)
Matchings may not be unique
Matching in Directed Networks
Matching: a set of directed edges without common heads and common tails
Matched node: the head node of a matching edge
Minimum Inputs Theorem
Q: How many?
A: The minimum number of inputs ND needed is:
Case 1: If there is a perfect matching, then ND = 1
Case 2: If there is no perfect matching, then ND = number of unmatched nodes
Q: Where to put them? A: Case 1: Anywhere
Case 2: At unmatched nodes
Y. Y. Liu, J. J. Slotine, A. L. Barabási, Nature (2011)
Examples
Perfect Matching
1 input; anywhere
A Network of MIMO LTI Systems
xi
xi Rn
xi
Axi
Bui
yi Rm
ui Rp
yi Cxi
N ijHyj, yi Cxi, i1,2, ,N
Node system Networked system
Networked system with external control
j1
Axi
Some notations
xi i 1:
j1
i 0:
without external control
N ijHCxj iBui, i1,2, ,N with external control
Axi
Node system (A,B,C) Network structure
L [ij ] RNN
diag(1, ,
Coupling matrix H
External control inputs
N
)
Some counter-intuitive examples
Network structure
0 1 L111
Node system
Networked MIMO system
H 1 0
10
A1 0 B0
structurally controllable
C01
(A,B) is uncontrollable (A,C) is observable
state controllable
Network structure
Node system
Networked MIMO system
Some counter-intuitive examples
H 0
01
10 A B1
L
structurally controllable
11 0
10
C01
(A,B) is controllable (A,C) is observable
state uncontrollable
1
Now, in general …
A Network of MIMO LTI Systems
xi
j1
N
HyBu, yCx, i1,2, ,N
Axi
i 1:withexternalcontrol i 0:withoutexternalcontrol
ij j i i i i
Question:
Given the MIMO node system (A, B, C), network structure L, and coupling matrix H, how to determine external inputs Δ to guarantee the controllability?
A Network of MIMO LTI Systems
Ns
xi Axi ijHCxj ikBuk,
xiRn, i1, N ukRp,k1, s
ylRq, l1, r Necessary and Sufficient Condition
j1 k1 yl mDx
N
j1
lj j
L [ij ]RNN
[ij ]RNs
Nn SMoalutrtiixoenqXuatioRn of
Controllability
TT
XB 0, L XHC X ( I A)
If and only if
Has a unique solution X = 0 is X 0.
L. Wang, X.F. Wang, G.R. Chen and W.K.S. Tang, Automatica, 2016
Pinning Control of MIMO Networks
BREAK
10 minutes
Consensus and Control over Complex Networks
Swarm Dynamics / Modeling Consensus Protocols/Analysis Flocking Algorithms / Control DEMO
Fish Schooling
DEMO
Birds Flocking
Flocking: to congregate or travel in flock DEMO
Consensus
A position reached by a group as a whole
Battle space management scenario
Fireflies Synchronization
Attitude Alignment
The attitude of each spacecraft is synchronized with its two adjacent neighbors via a bi-directional communication channel
Swarming
Flocking
Rendezvous Consensus Synchrony Cooperation ……
Distributed coordination of
a network of agents:
Agents
Network
Distributed local control Global consensus
What are in common ?
Flocking
Flocking (Some Real Photos)
DEMO
Position:
xi (t 1) xi (t) vt Heading:
(t1) (t) R Vicsek, et al, Phys. Rev. Lett. (1995)
Vicsek Model
(t1) (t) R
Vicsek Model
DEMO
Stability Analysis
A group of M globally nonlinearly coupled individuals:
dx M
i g(xi xj), i1,2,…,M
dt j1
Attraction/Repulsion function:
||y||2 g(y)yabexp c
All the agents will converge to a spherical region: Bx:||xx|| x1M x b c/2exp(1/2)
Mi
i1 a
Gazi, Passino, IEEE Trans. Auto. Control, 2003, 48: 692–697
A/R function:
g(y) y[ga (|| y ||) gr (|| y ||)]
Linear attraction
Bounded repulsion
gr(||xi xj ||)||xi xj ||b,b0 xi(t)B (x(t))
ga(||xi xj ||)a,a0 ||y(y)||
Stability Analysis
xi (xi) M g(xi xj),i1,…,M
j1,ji
xi
M
Gazi, Passino, IEEE Trans. Sys. Man Cybern., B: 2004, 34(1): 539-557
b/a
M 1b 2 aM M
(t 1) F(t) lim(t) ss1
Convergence Analysis
t
F random matrix
ss steady state T
1 [1,1,…,1]
Contraction Mapping Principle
Jadbabaie, Lin, Morse, IEEE Trans. Auto. Control, 2003, 48(6): 988-1001 Moreau, IEEE Trans. Auto. Control, 2005, 50(2): 169-182
Consensus
Design a network connection topology, or design local control law, so that || xi x j || 0 (here, consensus = synchronization)
Consensus is achieved asymptotically if there exists an infinite sequence of bounded intervals such that the union of the graphs over such intervals is totally connected.
union
Olfati-Saber, Murray, IEEE Trans. Auto. Control, 2004, 49(9): 1520-1533
Small-World Networks
are better for Consensus
N / 2 Condition number — the smaller, the better Regular networks Small-world networks
1000 times average
Olfati-Saber, Amer. Control Conf. (2005)
Flocking
DEMO
Boids Flocking Model
Three Rules:
Separation: Steer to avoid crowding local flockmates
Alignment: Steer to move toward the average heading of local flockmates Cohesion: Steer to move toward the average position of local flockmates
Reynolds, “Flocks, herd, and schools: A distributed behavioral model,” Computer Graphics, 1987, 21(4): 15-24
http://www.red3d.com/cwr/boids/
Flocking: Small-World Organization Small-world communication generates “pseudo-leaders”
who control their neighbors
H.T. Zhang, G.R. Chen, PhysCon, Germany (2007)
Flocking: Small-World Organization
2 11 00
2
Leader
Pseudo leader
Follower
-1
-2
-0.5 0 0.5 1 1.5 2
Initial position of the flock
-1
-2
-0.5 0 0.5 1 1.5 2
Flock position after 40 iterations
Flocking: DEMO
H.T. Zhang, G.R. Chen, PhysCon, Germany (2007)
Another Flocking Algorithm: DEMO
J. Lu and G. Chen (simulation)
Movies: Robot Fish
Acknowledgement: College of Engineering, Peking University
Movie 1: Coordination
Acknowledgement: College of Engineering, Peking University
Movie 2: Cooperation
Acknowledgement: College of Engineering, Peking University
End