Proceedings of the 19th World Congress

The International Federation of Automatic Control

Cape Town, South Africa. August 24-29, 2014

On Networked Evolutionary Games

Part 2: Dynamics and Control ?
Daizhan Cheng ∗ Fehuang He ∗∗ Tingting Xu ∗
∗
Key Laboratory of Systems and Control, Institute of Systems Science,
Academy of Mathematics and Systems Science, Chinese Academy of
Sciences, Beijing 100190, P.R.China
(e-mail: [email protected], [email protected])
∗∗
Institute of Astronautics, Harbin Institute of Technology, Harbin
150080, P.R.China (e-mail: [email protected])

Abstract: As Part 2 of the paper “on networked evolutionary games”, this paper uses the
framework presented in Part 1 (Qi et al., 2014) to explore further issues about networked
evolutionary games (NEGs). First, the strategy profile dynamics (SPD) is constructed from the
fundamental evolutionary equations (FEEs). Using SPD, the control of NEGs are investigated.
Detailed mathematical models are obtained for both deterministic and dynamic cases respec-
tively. Then certain more complicated NEGs are explored. They are: (i) NEG with strategies
of different length information, which allows some players use longer history information such
as the information at t and t − 1 or so; (ii) NEG with Multi-Species, which allows an NEG
with various kinds of players, they play several different fundamental network games according
to their identities. (iii) NEG with time-varying payoffs. Since payoffs determine the evolution,
the network profile dynamics will be a time-varying one. These more complicated NEGs can
cover more general evolutions and they generalized the method proposed in Cheng et al.
(Preprint2013).

Keywords: Networked evolutionary game, fundamental evolutionary equation, network profile

dynamics, heterogeneous NEG, semi-tensor product of matrices

1. INTRODUCTION investigated. A detailed mathematical framework is pre-

sented in Section 3 as a standard k-valued logical control
In Part 1 of this paper an NEG is defined as following, networks. Then all the techniques for the control of k-
which was firstly proposed in Cheng et al. (Preprint2013). valued logical networks can be used. Section 4 considers
Definition 1. An NEG, game, denoted by ((N, E), G, Π), the NEGs where players can use different length of histor-
consists of three ingredients as: ical information to update their strategies. In Section 5 we
consider the NEGs with multi-species. That is, the players
(i) a network (graph) (N, E); are classified into several species, and players of different
(ii) a fundamental network game (FNG), G, such that if species play different roles in the networked games. Section
(i, j) ∈ E, then i and j play the FNG with strategies 6 considers when the fundamental network game has time-
xi (t) and xj (t) respectively. varying payoff functions. Section 7 is a brief conclusion.
(iii) a local information based strategy updating rule
(SUR).
2. FROM FEE TO NPD
It was proved that the fundamental evolutionary equation
(FEE) for each player can be obtained as The NPD is used to describe the evolution of the overall
xi (t + 1) = fi ({xk (t)|k ∈ U2 (i)}) , i = 1, · · · , n. (1) networked games. This section consider how to construct
the NPD of an NEG using its nodes’ FEEs. We consider
Then the network profile dynamics is uniquely determined two cases: (i) the FEEs are deterministic model; (ii) the
by FEEs. FEEs are probabilistic model.
We refer to Qi et al. (2014) and Cheng et al. (Preprint2013)
for details. 2.1 Deterministic Model
Part 2 of the paper considers several advanced problems
about NEGs. In Section 2 the SPD is constructed from Assume
FEEs. Using SPD, the control problems of NEGs are 
x1 (t + 1) = M1 x(t),

? This work was supported in part by National Natural Science .. (2)
.
Foundation (NNSF) of China under Grants 61074114, 61273013, and 

61333001. xn (t + 1) = Mn x(t),

978-3-902823-62-5/2014 © IFAC 281

19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

where x(t) = nni=1 xi (t) and Mi ∈ Lk×kn . Then we have α = 3/ 41 + 4/ 14 + 7/ 14 + 8/ 41 ,

the NPD as β = 3/ 21 + 7/ 12 ,
x(t + 1) = M x(t), (3) γ = 8/ 31 + 16/ 23 ,
where µ = 1/ 32 + 9/ 13 ,
M = M1 ∗ M2 ∗ · · · ∗ Mn ∈ Lkn ×kn . (4) λ = 18/ 61 + 20/ 16 + 26/ 13 + 28/ 13 ,
Example 2. Recall Example 12 of Part 1. We have p = 24/ 31 + 32/ 23 ,
x1 (t + 1) = Mf x4 (t)x5 (t)x1 (t)x2 (t)x3 (t) q = 26/ 21 + 28/ 12 ,
= Mf W[23 ,22 ] x(t) := M1 x(t), r = 27/ 14 + 28/ 14 + 31/ 14 + 32/ 14 ,
x2 (t + 1) = Mf x5 (t)x1 (t)x2 (t)x3 (t)x4 (t)
= Mf W[24 ,2] x(t) := M2 x(t), s = 29/ 21 + 31/ 12 .
x3 (t + 1) = Mf x(t) := M3 x(t),
x4 (t + 1) = Mf x2 (t)x3 (t)x4 (t)x5 (t)x1 (t) 3. MODELING CONTROLLED NEGS
= Mf W[2,24 ] x(t) := M4 x(t),
x5 (t + 1) = Mf x3 (t)x4 (t)x5 (t)x1 (t)x2 (t) Definition 4. Let ((N, E), G, Π) be an NEG, and N = U ∪
= Mf W[22 ,23 ] x(t) := M5 x(t). Z be a partition of N . We call ((U ∪ Z), E), G, Π) a
Finally, we have the NPD as controlled NEG, if the strategies of u ∈ U can be chosen
arbitrarily. As a result, z ∈ Z is called a state and u ∈ U
x(t + 1) = M x(t), (5) is called a control.
where
Using FEE, the strategy evolutionary equations can be
M = M1 ∗ M2 ∗ M3 ∗ M4 ∗ M5 expressed as (Cheng et al., Preprint2013)
= δ32 [1, 20, 8, 24, 15, 32, 16, 32, 29, 32, 32, 32, 31, 32, 32, 32 (6)
26, 28, 32, 32, 32, 32, 32, 32, 30, 32, 32, 32, 32, 32, 32, 32]. xi (t + 1) = Mi x(t), i = 1, · · · , n, (11)
where x(t) = nnj=1 xj (t). Assume U = {i1 , · · · , iq } with
1 ≤ i1 < i2 < · · · < iq ≤ n, and Z = {j1 , j2 , · · · , jp } with
2.2 Probabilistic Model 1 ≤ j1 < j2 < · · · < jp ≤ n, where p + q = n. Define
ur = xir , r = 1, · · · , q, and zs = xjs , s = 1, · · · , p.
Assume the strategies have the probabilistic k-valued We consider the deterministic case and the probabilistic
logical form as case separately.
xi (t + 1) = M1j x(t), with P r = pji , (7) (1) (Deterministic Case) Assume Mi ∈ Lk×kn . Then we
j = 1, · · · , si ; i = 1, · · · , n. have
Then we have zs (t + 1) = xjs (t + 1) = Mjs nni=1 xi (t)
= Mjs W[k,kiq −1 ] uq (t)x1 (t) n x2 (t) n · · ·
x(t + 1) = M x(t), (8)
x̂iq n · · · n xn (t)
where M ∈ Υkn ×kn can be calculated as = Mjs W[k,kiq −1 ] W[k,kiq−1 ] uq−1 (t)uq (t)
x1 (t) n x2 (t) n · · · n x̂iq−1 n · · ·
s2
s1 X sn
" n ! #
X X Y j j1 j2 jn
M= ··· p i M1 ∗ M2 ∗ · · · ∗ M n .
i
x̂iq n · · · n xn (t)
j1 =1 j2 =1 jn =1 i=1 = ···
(9) = Mjs n1r=m W[k,kir +m−r−1 ] u(t)z(t),
where u(t) = nqi=1 ui (t), and z(t) = npi=1 zi (t). The
We use an example to depict it.
notation x̂s means this factor is removed.
Example 3. Recall Example 13 of Part 1. In fact, we can Define
use Table 5 there to calculate M row by row. For instance,
Ψs := Mjs n1r=m W[k,kir +m−r−1 ] ∈ Lk×kn , (12)
it is obvious that
1 then we have
Col1 (M ) = Col2 (M ) = Col3 (M ) = δ32 .
3 4
zs (t + 1) = Ψs u(t)z(t), s = 1, · · · , p. (13)
As for Col4 (M ), with probability 1/4 it could be δ32 or δ32
7 8 Set
or δ32 or δ32 . That is,
Ψ := Ψ1 ∗ Ψ2 ∗ · · · ∗ Ψp ∈ Lkp ×kn . (14)
Col4 (M ) = [0, 0, 14 , 14 , 0, 0, 41 , 41 , 0, 0, 0, 0, 0, 0, 0, 0,
The controlled network profile evolutionary equation
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]T .
is expressed as
We simply express it as z(t + 1) = Ψu(t)z(t). (15)
δ32 3/ 14 + 4/ 41 + 7/ 41 + 8/ 14 .
 
This is a standard k-valued logical control network.
Using this notation and a similar computation, we have (2) (Probabilistic Case) Assume
M = δ32 [1, 1, 1, α, 1, α, β, γ, µ, λ, 11, 32, λ, 32, 32, 32 Mi = Miji ∈ Lk×kn , with P r = pji i (16)
1, 1, 1, α, 1, 22, α, p, µ, q, r, 32, s, 32, 32, 32], j = 1, · · · , ri , i = 1, · · · , n.
(10) Then for each choice: {j1 , · · · , jn 1 ≤ ji ≤ ri } we
where can use {Miji i = 1, · · · , n} to construct Ψj1 ,··· ,jn ,

282
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

using the technique developed for deterministic case. where L can be calculated by using the technique proposed
Finally we have (15) again with in Example 17 in Part 1 (Qi et al., 2014) as
r1 X r2 rn Yn
X X L = δ8 [1, 1, 1/ 32 + 3/ 13 , 6/ 31 + 7/ 23 , 1, 1, 5/ 23 + 7/ 13 , 8, 1,
Ψ= ··· pji i Ψj1 ,··· ,jn ∈ Υkp ×kn . (17)
j1 =1 j2 =1 jn =1 i=1
1/ 21 + 2/ 21 , 3, 8, 1, 1/ 12 + 2/ 12 , 7/ 21 + 8/ 12 , 8,
1, 1/ 21 + 2/ 21 , 5/ 13 + 7/ 32 , 8, 1, 6, 7, 8,
Note that a general procedure is provided above. But for
a particular NEG, the process may be simplified. We use 1, 2/ 31 + 4/ 32 , 8, 8, 1/ 12 + 2/ 12 , 7/ 31 + 8/ 23 , 8, 8].
some examples to depict this.
Example 5. Recall Example 12 in Part 1. Assume players 4. NEG WITH STRATEGIES OF DIFFERENT
2 and 4 are controls and the others are states. That is, LENGTH INFORMATION
u1 = x2 , u2 = x4 , z1 = x1 , z2 = x3 , z3 = x5 .
Then we have Definition 7. Given an NEG ((N, E), G, Π).
x1 (t + 1) = Mf x4 (t)x5 (t)x1 (t)x2 (t)x3 (t) (i) A player, say, i, is said to use length-r (historic)
= Mf W[23 ,22 ] x1 (t)x2 (t)x3 (t)x4 (t)x5 (t) information, if
= Mf W[23 ,22 ] W[2,23 ] x4 (t)x1 (t)x2 (t)x3 (t)x5 (t)
= Mf W[23 ,22 ] W[2,23 ] W[2,22 ] u1 (t)u2 (t)z1 (t)z2 (t)z3 (t) xi (t + 1) = fi ({xj (t), xj (t − 1), · · · , xj (`),

(21)
:= L1 u(t)z(t), cj (t), · · · , cj (`) j ∈ U (i)} ,
where L1 = Mf W[23 ,22 ] W[2,23 ] W[2,22 ] , u(t) = u1 (t)u2 (t), where ` = max{0, t − r + 1}.
z(t) = z1 (t)z2 (t)z3 (t). Similarly, we can have (ii) The NEG is said to have strategies of different length
zi (t + 1) = Li u(t)z(t), i = 1, 2, 3, (18) (historic) information, if there is a partition N = N1 ∪
N2 ∪ · · · ∪ Ns , Ni ∩ Nj = ∅ (i 6= j) such that a player
where j ∈ Nr implies that j is with r-length information.
L1 = Mf W[23 ,22 ] W[2,23 ] W[2,22 ]
= δ2 [1, 2, 1, 2, 2, 2, 2, 2, 1, 2, 1, 2, 2, 2, 2, 2, Now assume i uses length-r information and let t ≥ r − 1.
2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2]; Then we have
L2 = Mf W[2,23 ] W[2,22 ] xi (t + 1)
= δ2 [1, 1, 2, 2, 1, 1, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, =fi ({xj (t), xj (t − 1), · · · , xj (t − r + 1),
2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2]; cj (t), · · · , cj (t − r + 1) j ∈ U (i)}


L3 = Mf W[22 ,23 ] W[2,23 ] W[2,22 ]

=fi {xj (t), xj (t − 1), · · · , xj (t − r + 1) j ∈ U2 (i)} .
= δ2 [1, 2, 1, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, (22)
1, 2, 1, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2].
Finally, we have controlled NEG as Note that in the above equation the fi in the first equality
is different from the fi in the second equation. To avoid
z(t + 1) = Lu(t)z(t), (19) notational complexity, we use the same notation. Now for
where each j we define
L = L1 ∗ L2 ∗ L3 z1j (t + 1) := xj (t)
= δ8 [1, 6, 3, 8, 6, 6, 8, 8, 4, 8, 4, 8, 8, 8, 8, 8,
7, 8, 7, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8]. z2j (t + 1) := z1j (t) = xj (t − 1)
.. (23)
Example 6. Recall Example 13 in Part 1. Assume players .
j j
3 and 4 are controls and the others are states. That is, zr−1 (t + 1) := zr−2 (t) = xj (t − r + 2).
u1 = x3 , u2 = x4 , z1 = x1 , z2 = x2 , z3 = x5 . Using this set of new variables, we can express (22) into a
Then we have normal form as
2
 j
L1 = M1 W[2,2 3] 
 z1 (t + 1) = xj (t)
z2j (t + 1) = z1j (t)

= δ2 [1, 1, 1, 2, 1, 1, 2, 2, 1, 1, 1, 2, 1, 1, 2, 2,



 ..

1, 1, 2, 2, 1, 2, 2, 2, 1, 1, 2, 2, 1, 2, 2, 2];

.
2
L2 = M2 W[2,2 3] j
zr−1
 j
(t + 1) = zr−2 (t), j ∈ U2 (i)

= δ2 [1, 1, 1/ 32 + 2/ 13 , 1/ 31 + 2/ 23 , 1, 1, 1/ 32 + 2/ 31 , 2,






1, 1, 2, 2, 1, 1, 2, 2, 1, 1, 1/ 13 + 2/ 23 , 2, 1, 1, xi (t + 1) = fi {xj (t),z1j (t),z2j (t),··· ,zr−1
j
(t) j∈U2 (i)} .


2, 2, 1, 1/ 31 + 2/ 23 , 2, 2, 1, , 1/ 31 + 2/ 23 , 2, 2]; (24)
2
L3 = M5 W[2,23]
Define
n o
= δ2 [1, 1, 1, 2, 1, 1, 1, 2, 1, 1/ 21 + 2/ 12 , 1, 2, 1,
j
y i = z1j , · · · , zr−1 j ∈ U2 (i) ∪ {xi },
1/ 12 + 2/ 21 , 1/ 21 + 2/ 21 , 2, 1, 1/ 21 + 2/ 12 , then we have
1, 2, 1, 2, 1, 2, 1, 2, 2, 2, 1/ 21 + 2/ 12 , 2, 2, 2]. y i (t + 1) = Fi {y j (t) j ∈ U2 (i)} .


(25)
Finally, we have the networked profile evolutionary equa-
tion as Then the technique developed in Part 1 for standard NEGs
is applicable for this case. Finally, we consider the initial
z(t + 1) = Lu(t)z(t), (20) values. We consider {xi (0), · · · , xi (r − 1) i = 1, · · · , n} as

283
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

the initial values. In fact, only {xi (0) i = 1, · · · , n} are y1 (t + 1) = x1 (t)

5
real initial values. Then we can use the following equation = Dr[2 ,2] z(t)x1 (t)
5 [2,25 ]


xi (1) = Fi {xj (0) j ∈ U2 (i)}
= Dr[2 ,2] z(t)D f x(t)
xi (2) = Fi {xj (0), xj (1) j ∈ U2 (i)} [2,25 ]
 5
= Dr[2 ,2] I25 ⊗ Df z(t)x(t)
..
.
:= M1 y(t),
xi (r − 1) = Fi {xj (0), xj (1), · · · , xj (r − 2) j ∈ U2 (i)} where
[2,25 ]
5
 
(26) M1 = Dr[2 ,2]
I25 ⊗ Df .
and the method similar to (23)–(25) to calculate all other Similarly, we can calculate that
initial values.
yi (t + 1) = Mi y(t), i = 2, · · · , 11,
To calculate the network profile dynamics of this kind where
of networks, we need the following lemma (Cheng et al., 5
 
[22 ,24 ]
Preprint2013) M2 = Dr[2 ,2]
I25 ⊗ Dr[2,2] Df
[23 ,23 ]
5
 2

Lemma 8. Assume X ∈ ∆p and Y ∈ ∆q . We define two M3 = Dr[2 ,2]
I25 ⊗ Dr[2,2 ] Df
dummy matrices, named by “front-maintaining operator” 5
 4 [25 ,2]

(FMO) and “rear-maintaining operator”(RMO) respec- M4 = Dr[2 ,2]
I25 ⊗ Dr[2,2 ] Df
 
tively, as: M5
5
= Dr[2 ,2] I25 ⊗ Dr[2,2 ]
5

[p,q]
Df | ·{z
= δp [1 · · 1} 2| ·{z
· · 2} · · · p · · · p], 5 5 6
(
Mf Dr[2,2 ] W[23 ,23 ] Dr[2 ,2 ] := M61 , P r = 0.8
q q
| {z }
q
M6 = [25 ,26 ]
Mf W[23 ,22 ] Df := M62 , P r = 0.2
Dr[p,q] = δq [1 2 · · · q 1 2 · · · q · · · 1 2 · · · q ]. [2,25 ] [25 ,26 ]
M7 = Mf D r W[24 ,22 ] Dr
[25 ,2] [25 ,26 ]
| {z } | {z } | {z }
| {z } M8 = M f Df Dr
p
[26 ,25 ]
Then we have M9 = Mf Dr
[25 ,2] 5 6

Df
[p,q]
XY = X. (27) M10 = Mf Df W[22 ,24 ] Dr[2 ,2 ]
[25 ,2] 5
,26 ]
M11 = Mf Df W[23 ,23 ] Dr[2 .
Dr[p,q] XY = Y. (28)
Then
We give an example to depict this. y(t + 1) = Ly(t),
Example 9. Consider an NEG ((N, E), G, Π), where the where
1
n
graph is a cycle of n = 6 nodes, and the FNG is the same L=
L1 := M1 ∗M2 ∗M3 ∗M4 ∗M5 ∗M6 ∗M7 ∗M8 ∗M9 ∗M10 ∗M11 ,
2
P r = 0.8,
L2 := M1 ∗M2 ∗M3 ∗M4 ∗M5 ∗M6 ∗M7 ∗M8 ∗M9 ∗M10 ∗M11 , P r = 0.2.
as in Example 12 of Part 1. Assume players 2, 3, 4, 5, 6
use length-1 information, and the SUR, Π, is the same as As for the initial value, we have
in Example 13 in Part 1; and the player 1 uses length-2 y(1) = (x1 (0), x2 (0), x3 (0), x5 (0), x6 (0),
information; and the SUR for t = 1 it is Π, for t > 1 the (30)
(x1 (1), x2 (1), x3 (1), x4 (1), x5 (1), x6 (1))),
SUR for player 1 is as follows: using Π to get xj ∗ (t) (t) and
with (x1 (0), x2 (0), x3 (0), x4 (0), x5 (0), x6 (0)) are free val-
xj ∗ (t−1) (t). Then we assume
 ues, and
x1 (t + 1) =
xj ∗ (t) (t), P r = 0.8, x(1) = M 0 x(0), (31)
xj ∗ (t−1) (t), P r = 0.2. where
Then the strategy dynamics for players 2, 3, 4, 5, 6 M 0 = M10 ∗ M20 ∗ M30 ∗ M40 ∗ M50 ∗ M60
are the same as in Example 12 of Part 1. The strategy [25 ,2] [25 ,2]
= (Mf Df W[24 ,22 ] ) ∗ (Mf Df W[25 ,2] )
dynamics for player 1 is [25 ,2] 5
∗(Mf Df ) ∗ (Mf Dr[2,2 ] W[2,25 ] )
z1 (t + 1) = x1 (t) [2,25 ] 5
z2 (t + 1) = x2 (t) ∗(Mf Dr W[22 ,24 ] ) ∗ (Mf Dr[2,2 ] W[23 ,23 ] ).
z3 (t + 1) = x3 (t)
z4 (t + 1) = x5 (t) Finally, we have
z5 (t + 1) = 
x6 (t)
f1 (x5 (t),x6 (t),x1 (t),x2 (t),x3 (t)), P r = 0.8 L = 0.8L1 + 0.2L2
x1 (t + 1) =
(f1 (z4 (t),z5 (t),z1 (t),z2 (t),z3 (t)), P r = 0.2
5
= δ2048 [1, 68/0.2 + 100/0.8, 136, 200/0.2 + 232/0.8, 15,
Mf Dr[2,2 ] W[23 ,23 ] x(t), P r = 0.8 80/0.2 + 112/0.8, 144, 208/0.2 + 240/0.8, 285,
=
Mf W[23 ,22 ] z(t), P r = 0.2.
. . . , 1984, 2048].
(29)
where x(t) = n6i=1 xi (t), z(t) = n5i=1 zi (t), Mf is the same Then, after 18 times iterations L converges to the following
as in Example 13 in Part 1. Denoted by matrix
 δ2048 [1, 2048, 2048, . . . , 2048].
z (t), i = 1, · · · , 5;
yi (t) = i According to this matrix (the whole set of data is omitted),
xi−5 (t), i = 6, · · · , 11. we splite ∆2048 into three subsets:
Then D1 = δ2048 {1, 129, 257, 385};

284
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

D2 = δ2048 {65 193 321 449 513 577 641 Table 1. Payoffs → Dynamics (Example 11)
705 769 833 897 961 1025 1089
Profile 11111 11112 11121 11122 11211 11212 11221 11222
1153 1217 1281 1345 1409 1473 1537 c1 -1 -1 -1 -1 -1 -1 -1 -1
1601 1665 1729 1793 1857 1921 1985}; c2 5/4 3/4 3/4 1/4 1/4 1/4 1/4 -1/4
c3 1 1 1 1 0 0 0 0
D3 = ∆2048 \ (D1 ∪ D2 ) . c4 1 1 0 0 1 1 0 0
1
If initial state x0 ∈ D1 , then x(t) → δ2048 as t → ∞. Else f1 1 1 1 1 1 1 1 1
1 2048 f3 1 1 1 1 1 1 1 2
if x0 ∈ D2 , then x(t) → 0.8 ∗ δ2048 + 0.2 ∗ δ2048 as t → ∞. f4 1 1 1 1 1 1 1 2
2048
Else where x0 ∈ D3 , then x(t)δ2048 as t → ∞. ..
.
Note that not all x ∈ ∆2048 can be chosen as the initial Profile 22111 22112 22121 22122 22211 22212 22221 22222
c1 1 1 1 1 1 1 1 1
value, because the initial value should satisfy (30)–(31).
c2 1/4 1/2 1/2 3/4 1/2 3/4 3/4 1
c3 0 0 0 0 2 2 2 2
c4 0 0 2 2 0 0 2 2
5. NEG WITH MULTI-SPECIES f1 2 2 2 2 2 2 2 2
f3 2 2 2 2 2 2 2 2
f4 2 2 2 2 2 2 2 2
Definition 10. An NEG is said to have s species, if there
is a partition N = N1 ∪ N2 ∪ · · · ∪ Ns , a set of fundamental Table 2. Payoffs → Dynamics (Example 11)
games {Gi,j 1 ≤ i, j ≤ s}
Profile 111111 111112 111121 111122 111211 111212 111221 111222
c1 -1 -1 -1 -1 -1 -1 -1 -1
To avoid the notational complexity, we assume s = 2. We c2 5/4 5/4 3/4 3/4 3/4 3/4 1/4 1/4
c3 1 1 1 1 1 1 1 1
call these two kinds of players (nodes) white (W ) and black c4 1 1 1 1 0 0 0 0
(B) respectively. Then there are three different NEGs: Gw , c5
c6
0
-1
0
2
1
-1
0
0
0
-1
0
2
1
-1
0
0
Gb , and Gm . It is reasonable to assume that Gw and Gb , f2 1 1 (2/3,1/3) 1 1 1 (1/2,1/2) 1
f5 1 2 2 1 1 2 2 1
which are the games between two white and two black ..
.
players respectively, are symmetric, and Gm , which is the Profile 122111 122112 122121 122122 122211 122212 122221 122222
game between a white and a black players, is asymmetric. c1
c2
2
1/4
2
1/4
2
1/2
2
1/2
2
1/2
2
1/2
2
3/4
2
3/4
Assume in all the three games there are k strategies for c3
c4
2
0
2
0
2
0
2
0
2
2
2
2
2
2
2
2
each player. Then each player has k ×k possible strategies, c5 -1/2 -1/2 2 1 -1/2 -1/2 2 1
c6 -1 2 -1 0 -1 2 -1 0
denoted by zi (t) = xi (t) n yi (t), where xi is the strategy f2 (1/2,1/2) (1/2,1/2) (1/3,2/3) (1/2,1/2) (1/3,2/3) (1/3,2/3) (1/4,3/4) (1/3,2/3)
against white neighbors and yi is the strategy against black f5 2 2 2 2 2 2 2 2

neighbors. We give an example to depict this. Table 3. Payoffs → Dynamics (Example 11)
Example 11. A game with its graph depicted in Fig. 1,
Profile 211111 211112 211121 211122 211211 211212 211221 211222
where 4 nodes are white and 2 others are black. Assume c1 0 0 0 0 0 0 0 0
k = 2 and the payoff bi-matrices for three FNGs are c2 2 2 3/2 3/2 3/2 3/2 1 1
c3 1 1 1 1 1 1 1 1
described by c4 1 1 1 1 0 0 0 0
c5 0 0 1 0 0 0 1 0
(i) Gw is S-2 with parameters as (Snowdrift Game) c6 -1 2 -1 0 -1 2 -1 0
f2 1 1 1 1 1 1 (2/3,1/3) 1
R = −1; S = −1; T = 2; P = 0; f5 1 (1/2,1/2) 1 1 1 2 (1/2,1/2) 1
..
(ii) Gb is S-2 with parameters as (Hawk-Dove Game) .
Profile 222111 222112 222121 222122 222211 222212 222221 222222
R = −1; S = 2; T = 0; P = 1; c1
c2
1
1/2
1
1/2
1
3/4
1
3/4
1
3/4
1
3/4
1
1
1
1
c3 2 2 2 2 2 2 2 2
(iii) Gm is A-2 with parameters as c4 0 0 0 0 2 2 2 2
c5 -1/2 -1/2 2 1 -1/2 -1/2 2 1
A = 2, B = 1, C = 0, D = 0, E = 0, F = 0, G = 1, H = 2. c6 -1 2 -1 0 -1 2 -1 0
f2 2 2 2 2 2 2 2 2
f5 2 2 2 2 2 2 2 2
1
Table 4. Payoffs → Dynamics (Example 11)

2 Profile 111 112 121 122 211 212 221 222

c5 0 3/2 -1/2 0 -1/2 1 1/2 1
3 4 c6 -1 2 -1 0 -1 2 -1 0
f6 1 2 2 2 1 2 2 2

5 where
L = δ64 [1, 4, 4, 2, 1, 4, 4, 2, 1, 4, 4, 2, 1, 4, 32, 32, 15, 16,
16, 16, 15, 16, 16, 16, 15, 16, 16, 16, 15, 16, 16, 16,
1, 2, 2, 2, 1, 4, 2, 2, 1, 4, 2, 2, 1, 4, 20, 2, 63, 64,
6 64, 64, 63, 64, 64, 64, 63, 64, 64, 64, 63, 64, 64, 64].
Fig. 1. Graph for Example 11 This NEG has two fixed points (1, 1, 1, 1, 1, 1) and (2, 2, 2,
2, 2, 2). Besides, it has two cycles with length 2, i.e.,
We can calculate fi as in Tables 1-4. (2, 2, 2, 2, 2, 1) → (2, 2, 2, 2, 1, 1) → (2, 2, 2, 2, 2, 1)
Using SUR UI-1, We have the NEG as and
x(t + 1) = L n6i=1 xi (2, 2, 1, 1, 1, 1) → (2, 1, 1, 1, 1, 1) → (2, 2, 1, 1, 1, 1).

285
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

6. NEG WITH TIME-VARYING PAYOFFS (iv) When (R, T ) ∈ V I, the FEE (32) is specified as f4
where
Definition 12. An NEG is said to have varying payoffs, if α = 2; β = 2; γ = 2; θ = 2.
the parameters in the payoff bi-matrix of the NEG are
(v) When (R, T ) ∈ V , the FEE (32) is specified as f5
time-varying.
where
Example 13. Recall Example 12 of Part 1, where network
α = 2; β = 1; γ = 1; θ = 2.
graph is Fig. 1 (a) and the SUR is UI-1. As for the FNG,
we let the non-zero parameters be flexible, that is: FNG is Finally, we consider the NEG with time-varying payoff
S-2 with constrains: R = S, P = 0 (the generalized game parameters as
R = sin π t
  
of Snowdrift).
6 
Similar to Example 12 of Part 1, the FEE can be de- T = cos π t .
termined via Table 5. The parameters in Table 5 are as 6
follows: Then it is clear that the FEE becomes a periodic function
1, R ≥ T2 with period 12, precisely,

1, R ≥ T
α= β= xi (t + 1) = f (t) ni+2 ∀i,
2, R < T, 2, R < T2 , j=i−2 xi (t), (33)

1, R ≥ max{ T2 , 0}

1, R ≥ max{T, T2 } where
γ= θ= f1 , t ∈ {12m,12m+1,12m+9,12m+10,12m+11}

2, otherwise, 2, otherwise. 

f3 , t = 12m + 8

f (t) =
Table 5. (Parameter-Depending) Payoffs → f4 , t ∈ {12m+2,12m+4,12m+5,12m+6,12m+7}

Dynamics 
f5 , t = 12m + 2, m ∈ Z+ .
Profile 11111 11112 11121 11122 11211 11212 11221 11222
f 1 1 α β α α β γ Note that this is correct for Sn and R∞ (Graph of R with
Profile 12111 12112 12121 12122 12211 12212 12221 12222 all integers as nodes).
f α α α θ β β 2 2
Profile 21111 21112 21121 21122 21211 21212 21221 21222 T
f 1 1 α β α α β γ
Profile 22111 22112 22121 22122 22211 22212 22221 22222
f β β θ β γ γ 2 2

Hence, the FEE is IV V

f = δ2 [1, 1, α, β, α, α, β, γ, α, α, α, θ, β, β, 2, 2,
1, 1, α, β, α, α, β, γ, β, β, θ, β, γ, γ, 2, 2] ni+2
j=i−2 xj .
θ
(32) O R
Denote by I
T
A = {R
 ≥ T }; T  ≥ 2 };
B = {R III
C = R ≥ max{ 2 , 0} ; D = R ≥ max{T, T2 } . II
Let Θ (with 0 ≤ Θ < 2π) be defined by
R T
sin(Θ) = √ ; cos(Θ) = √ .
2
R +T 2 R + T2
2
Fig. 2. A Partition of Parameter Space
Then the parameter space can be decomposed into 5 parts
as shown in Fig. 2, where 7. CONCLUSION
π 3
n o n o
I = A∪B∪C∪D = 0≤Θ≤ ∪ π ≤ Θ < 2π
4 2 Based on the FEEs of all players, the SPD of overall
3
n o
II = A ∪ B ∪ C c ∪ D = π + arctan(2) ≤ Θ < π NEG is constructed. Using SPD, the controlled NEGs are
2
n
1
o introduced and converted into standard k-valued logical
III = A ∪ B c ∪ C c ∪ Dc = π + π ≤ Θ < π + arctan(2) control networks. Then some more complicated kinds of
4
1
n o
V I = Ac ∪ B c ∪ C c ∪ D c = arctan(2) < Θ < π + π NEGs are investigated. They are (i) players using different
4 o length historical information; (ii) players of multi-species;
1
n
V = Ac ∪ B ∪ C ∪ D c = π + π < Θ ≤ arctan(2) . and (iii) the fundamental network game with time-varying
4
payoffs. The formulations for these three more complicated
It follows that kinds of NEGs are obtained, some interesting results are
(i) When (R, T ) ∈ I, the FEE (32) is specified as f1 investigated.
where
REFERENCES
α = 1; β = 1; γ = 1; θ = 1.
(ii) When (R, T ) ∈ II, the FEE (32) is specified as f2 D. Cheng, F. He, H. Qi, T. Xu (Preprint2013). Mod-
where eling, analysis and control of networked evolutionary
games, http://lsc.amss.ac.cn/~dcheng/preprint/
α = 1; β = 1; γ = 2; θ = 1. NTGAME02.pdf (IEEE TAC, provisionally accepted).
(iii) When (R, T ) ∈ III, the FEE (32) is specified as f3 H. Qi, D. Cheng, H. Dong (2014). On Networked Evo-
where lutionary Games, Part 1: Formulation, Proc. of IFAC
α = 1; β = 2; γ = 2; θ = 2. World Congress 2014.

286