Game interactions and dynamics on networked populations

Transcription

1 Game interactions and dynamics on networked populations Chiara Mocenni & Dario Madeo Department of Information Engineering and Mathematics University of Siena (Italy) ({mocenni, Siena, May 3rd, 2016

2 Main idea behind Evolutionary game theory

3 Evolutionary game theory Infinite population of equal individuals (players or replicators) Each individual exhibits a certain phenotype (= it is preprogrammed to play a certain strategy). An individual asexually reproduces himself by replication. Offspring will inherit the strategy of parent. Nature favors the fittest Two replicators are randomly drawn from the population. Only one of them can reproduce himself. They play a game in which strategies are the phenotypes. Fitness of a replicator is evaluated like for game theory! The fittest is able to reproduce himself. The share of population with that phenotype will increase. Natural selection!

4 Replicator equation How to describe the evolution of the pure strategies distribution over the whole population? ẋ s = x s (p s (x) φ(x)) x s is the share of population with strategy s 0 x s 1, s xs = 1 t 0 p s(x) = es T Bx is the payoff obtained by a player with strategy s φ(x) = x T Bx is the average payoff obtained by a player B is the payoff matrix All Nash equilibria are rest points. A Lyapunov stable rest point is alway a Nash equilibrium. Weibull, J.W., Evolutionary Game Theory. MIT Press Hofbauer J., Sigmund K., Evolutionary games and population dynamics, Cambridge University Press

5 Replicator equation - some results Payoff matrices Rest points A NE B A dominates A NE B B dominates A NE NE NE B Bistability A NE B Coexistence Stable Unstable All "A" All "B" Some "A", some "B" Marginal stability and existence of infinite equilibria are also possible, e.g with payoff matrices [a b; a b].

6 Evolutionary games on networked populations...but in real world situations, things are different! Finite population of individuals Population has a structure: an individual can meet only his neighbors Individuals are described by vertices of a graph, edges are the connections between them. A = {a v,w } is the adjacency matrix of the graph In general, players aren t preprogrammed to play a pure strategy; they can be mixed (i.e. they can be partially cooperator and partially defector, selfish and altruist...) Efforts to extend the replicator equation to networks The replicator equation on graphs has been obtained for networks with infinite vertices Algorithmic approaches have been used to consider a finite population evolving in discrete time

7

8 Our proposal for a replicator equation on graphs Main idea A mixed player is an infinite population of replicators! Classical replicator equation Replicator equation on graph Replicator player Mixed player Ohtsuki, H., Nowak, M.A., The replicator equation on graphs. J. Theor. Biol., Vol. 243, N. 1, pp Szabó G., Fáth G., Evolutionary Games on Graphs. Physics Reports, Vol. 446, pp

9 Step 1. From the RE to the multipopulation RE a) The population is well mixed and the individuals differ only for the chosen pure strategy b) The population is splitted in subpopulations with different characteristics (not the strategies) c) The subpopulations are organized as the vertices of a fully connected graph

10 Modeling phase: preliminaries x v,s is the share of replicators in subpopulation v with strategy s x v = [x v,1... x v,m ] T is the strategy distribution of the subpopulation v p v,s = π v (e s, x v ) 1 is the average payoff earned by a replicator of subpopulation v with strategy s (e s is the s-th standard versor of R M ) φ v = π v (x v, x v ) = M s=1 x v,sp v,s is the average payoff earned by a replicator of subpopulation v 1 In general, π v (x v, x v ) is the payoff earned by the mixed player v when he plays strategy x v.

11 The dynamics of interactions (1 / 3) τ: interval between interactions n v (t): number of replicators inside subpopulation v p v,s (t): fitness of a replicator which uses s in subpopulation v at t p v,s (t)τ: number of offspings produced by each replicator n v (t)x v,s (t)p v,s (t)τ: number of offsprings produced by the replicators with strategy s in v n v (t)x v,s (t)(1 + p v,s (t)τ): number of replicators (parents and offsprings) with strategy s in v at t + τ n v (t + τ) = n v (t)(1 + φ v (t)τ): subpopulation size at t + τ

12 The dynamics of interactions (2 / 3) What happens to x v,s (t + τ)? Then: x v,s (t + τ) = n v (t)x v,s (t)(1 + p v,s (t)τ) n v (t + τ) = x v,s(t)(1 + p v,s (t)τ) 1 + φ v (t)τ x v,s (t + τ) x v,s (t) τ = x v,s(t)(p v,s (t) φ v (t)) 1 + φ v (t)τ When τ goes to 0... ẋ v,s (t) = x v,s (t)(p v,s (t) φ v (t))

13 The dynamics of interactions (3 / 3) ẋ v,s (t) = x v,s (t)(p v,s (t) φ v (t)) When the strategy s is fitter than the average, then the share of population that uses s grows; specifically, when p v,s (t) > φ v (t), then the variable x v,s (t) increases, and decreases otherwise. Multipopulation replicator equation Weibull, J.W., Evolutionary Game Theory. MIT Press

14 Step 2. The multipopulation RE on graphs The subpopulations are organized as the vertices of a generic finite graph and each vertex behaves as a population player The replicators of a vertex play 2-players games with the replicators of all the other vertices connected to him Calculate the payoffs of the vertex players

15 Step 3. Calculate the payoffs (1/3) Classical replicator equation Replicator equation on graph Replicator player Mixed player A vertex player sees all his neighbors as an average player This game is equivalent to a N-players M-strategies game

16 Step 3. Calculate the payoffs (2/3) The payoffs depend on the adjacency matrix A = {a v,w } of the graph. Payoff obtained by player v when players use strategies s 1,..., s N is a tensor: ( ) π v G (s 1,..., s N ) = e T 1 N s v B v a v,w e sw d v v=1 When strategy distributions x v are known for each vertex player, then the average payoff can be evaluated as follows: ( M M N ) πv G =... x w,sw π v G (s 1,..., s N ). s 1 =1 s N =1 w=1 Jackson M. O., Zenou Y., Games on Networks, Handbook of Game Theory, Vol. 4, Peyton Young and Shmuel Zamir, eds., Elsevier Science

17 Step 3. Calculate the payoffs (3/3) The extension to the mixed strategy set allows to evaluate the following payoff functions: Payoff obtained by player v with strategy s: p G v,s = e T s B v k v Average payoff obtained by player v: φ G v = x T v B v k v where k v = 1 d v N v=1 a v,w x w

18 Multipopulation RE + suitable payoffs = RE on graphs x v,s is the probability that player v will play strategy s the probability that a replicator inside v is preprogrammed to use strategy s. Then, the replicator equation on graphs (RE-G) is: or ẋ v,s = x v,s (p G v,s φ G v ) The Cauchy problem: ẋ v,s = x v,s (pv,s G φ G v ) v V, s S, x v,s (t = 0) = c v,s

19 Properties of RE-G: invariance and pure stationary points Simplex invariance x v (0) M v x v (t) M v, t > 0. Pure strategies are stationary points If x v = e s, then φ G v = p G v,s, ẋ v,s = x v,s (p G v,s φ G v ) = 0. Moreover, ẋ v,r = x v,r (p G v,r φ G v ) = 0, since x v,r = 0, for all r s. Then x v is a stationary point of the equation.

20 Properties of RE-G: Evolutionary stable strategies Evolutionary stable strategy profile - The idea of evolutionary stable strategy profile in the multipopulation context is slightly different from the standard case. Specifically, {x 1,..., x N } is an evolutionary stable strategy profile if, for every strategy profile {y 1,..., y N } {x 1,..., x N } there exists some ɛ y such that for all ɛ (0, ɛ y ), and with the strategy profile {z 1,..., z N }, defined for each v {1,..., N} as z v = ɛy v + (1 ɛ)x v, we have π w (x w, z w ) > π w (y w, z w ) for some w {1,..., N}. An evolutionary stable strategy profile {x 1,..., x N } is also strict Nash equilibrium of the underlying game, and viceversa.

21 Properties of RE-G: equivalence to the standard RE Theorem Let X(t) be the unique solution of the Cauchy problem RE-G, where x v,s (0) = c s v and B v = B. Moreover, let y(t) be the unique solution of the standard Cauchy problem with y s (0) = c s. Then, x v (t) = y(t) v, t 0.

22 Properties of RE-G: stationary points, NE and ESS Let {x 1,..., x N } be a stationary point of RE-G. If x v,s > 0 v, s, then {x 1,..., x N } is a Nash equilibrium of underlying game. If {x 1,..., x N } is Lyapunov stable stationary point of RE-G, then it it is a Nash equilibrium. Evolutionary stable strategy profiles are asymptotically stable stationary points of RE-G, and viceversa.

23 The RE-G for (N, 2)-games The payoff matrix for 2-strategies games is: [ ] bv,1,1 b B v = v,1,2. b v,2,1 b v,2,2 Let σ v,1 = b v,1,1 b v,2,1, σ v,2 = b v,2,2 b v,1,2 and k v (y) = [k v,1 (y) k v,2 (y)] T, then the RE-G becomes: ẏ v = y v (1 y v )f v (y), (1) where f v (y) = σ v,1 k v,1 (y) σ v,2 k v,2 (y). (2)

24 Stationary points of RE-G for (N, 2)-games The set of stationary points is: { } Θ = y [0, 1] N : v yv = 0 yv = 1 f v (y ) = 0. 1 Pure stationary points The set Θ p = {0, 1} N is a subset of Θ. Pure stationary points always exist, for every adjacency matrix A and payoff matrices B v. 2 Mixed stationary points The set Θ m = (0, 1) N Θ is a subset of Θ. Equivalently, Θ m = {y (0, 1) N : σ v,1 k v,1 (y ) = σ v,2 k v,2 (y ) v}. 3 Pure/Mixed stationary points The set Θ mp = Θ \ (Θ p Θ m ) is a subset of Θ.

25 (N, 2)-games: stability of pure stationary points (1/2) Theorem Let y Θ p. Then, the following statements are equivalent: (a) ((σ v,1 k v,1 (y ) σ v,2 k v,2 (y ) y v = 0) (σ v,1 k v,1 (y ) σ v,2 k v,2 (y ) y v = 1)) v (b) λ v (J(y )) 0 v (c) y Θ NE Corollary Let y Θ p. Then, the following statements are equivalent: (a ) ((σ v,1 k v,1 (y ) < σ v,2 k v,2 (y ) y v = 0) (σ v,1 k v,1 (y ) > σ v,2 k v,2 (y ) y v = 1)) v (b ) λ v (J(y )) < 0 v (c ) y Θ NES

26 (N, 2)-games: stability of pure stationary points (2/2) Previous theorems provide an interpretation of the stability of stationary points of RE-G. Indeed, σ v,s s = 1, 2, are the gains obtained when choosing pure strategy 1 or 2 k v,s (y ), s = 1, 2, are related to the number of players connected to vertex v When for all vertices vs, the gain of neighbor players which choose the same pure strategy as player v is bigger, then the stationary state is stable

27 (N, 2)-games: stability of mixed stationary points Theorem Suppose that σ v,1 = σ 1 and σ v,2 = σ 2 v with sign(σ 1 ) = sign(σ 2 ) 0. Then there always exists a stationary point y Θ m, such that yv = σ 2 σ 1 +σ 2 v. Moreover, λ(j(y )) = σ 1σ 2 σ 1 +σ 2 λ(a). If A is invertible, then the mixed stationary state y is unique and does not depend on A itself If A is not invertible, then there exist other stationary points which depend on A The linearized stability of y depends on the eigenvalues of the adjacency matrix A

28 Replicator equation on graphs: simulation results Bistable payoff matrix: B = [ ] Graph at t = 0 2 (0.99) Vertices dynamics (percentage of strategy A) Vertex 1 Vertex 2 3 (0.99) 6 (0.99) (0.99) 4 (0.99) 5 (0.99) t Graph at t = 30 2 (1.00) Strategies percentage over all the graph Strategy A Strategy B 3 (1.00) 6 (1.00) (1.00) 4 (1.00) 5 (1.00) t Since all vertices are the same at initial time, then evolution is equivalent to classical replicator equation.

29 Replicator equation on graphs: simulation results Bistable payoff matrix: B = [ ] Graph at t = 0 2 (0.01) Vertices dynamics (percentage of strategy A) Vertex 1 Vertex 2 3 (0.99) 6 (0.99) (0.99) 4 (0.99) 5 (0.99) t Graph at t = 30 2 (1.00) Strategies percentage over all the graph Strategy A Strategy B 3 (1.00) 6 (1.00) (1.00) 4 (1.00) 5 (1.00) t Network is resistant to the mutator.

30 Replicator equation on graphs: simulation results Bistable payoff matrix: B = [ ] Graph at t = 0 2 (0.99) Vertices dynamics (percentage of strategy A) Vertex 1 Vertex 2 3 (0.99) 6 (0.99) (0.01) 4 (0.99) 5 (0.99) t Graph at t = 30 2 (0.50) Strategies percentage over all the graph Strategy A Strategy B 3 (0.50) 6 (0.50) (0.50) 4 (0.50) 5 (0.50) t The mutator is stronger thanks to his connections. We have a diffusive process and coexistence of strategies

31 Replicator equation on graphs: simulation results Bistable payoff matrix: B = [ ] Graph at t = 0 2 (0.01) Vertices dynamics (percentage of strategy A) Vertex 1 Vertex 2 Vertex 3 3 (0.99) 6 (0.99) (0.01) 4 (0.99) 5 (0.99) t Graph at t = 30 2 ( 0.00) Strategies percentage over all the graph Strategy A Strategy B 3 ( 0.00) 6 ( 0.00) ( 0.00) 4 ( 0.00) 5 ( 0.00) t The central mutator is helped by another mutator. They impose their strategy to all the population!

32 Replicator equation on graphs: simulation results Prisoners dilemma game: B = [ ] Resilience of cooperation in non strict Prisoners dilemma

33 The attractive mixed Nash equilibrium of RE is not homogeneously stable in the RE-G Replicator equation on graphs: simulation results Attractive mixed Nash equilibrium: B = [ ]

34 Conclusions and Future Work The theoretical formulation of the RE-G provides a new way for describing the dynamical interactions among individuals in a society obtaining a general framework to derive suitable equations of such interactions understanding the properties of unexplained real world phenomena Future work will investigate complex graph topologies and their influence on the system dynamics: properties of the adjacency matrix A equivalence of the model with infinite dimensional systems when the number of vertices becomes high asymptotic stability of equilibria of the ordinary differential equations by means of nonlinear techniques Network reconstruction by solving suitable inverse problems