physics0411091/text.tex
1: \documentclass[aps,prl,preprint,floats,twocolumn,floatfix,english]{revtex4}
2: \usepackage[T1]{fontenc}
3: \usepackage[latin1]{inputenc}
4: \usepackage{amsmath}
5: \usepackage{setspace}
6: \usepackage{epsfig}
7: 
8: \makeatletter
9: \providecommand{\LyX}{L\kern-.1667em\lower.25em\hbox{Y}\kern-.125emX\@}
10: 
11: \usepackage{babel}
12: \makeatother
13: \begin{document}
14: \title{Spatial snowdrift game with myopic agents}
15: 
16: \author{Marko Sysi-Aho $^1$ }
17: \email[]{msysiaho@lce.hut.fi}
18: \author{Jari Saram\"{a}ki $^1$ }
19: \author{J\'{a}nos Kert\'{e}sz $^{1,2}$ }
20: \author{Kimmo Kaski $^{1}$ }
21: 
22: 
23: \affiliation{$^1$Laboratory of Computational Engineering, Helsinki
24: University of Technology, Espoo, Finland\\
25: $^2$Department of Theoretical Physics,
26:          Budapest University of Technology and Economics, Budapest, Hungary}
27: 
28: \begin{abstract}
29: We have studied a spatially extended 
30: snowdrift game, in which the players are located on the sites of 
31: two-dimensional square lattices and repeatedly have to choose one of the two strategies, either cooperation (C) or defection (D). 
32: A player interacts with its nearest neighbors only, and 
33: aims at playing a strategy which maximizes its instant pay-off, 
34: assuming that the neighboring agents retain their strategies. 
35: If a player is not content with its current strategy, it will change
36: it to the opposite one with probability $p$ next round. Here we show through simulations and analytical 
37: approach that these rules result in cooperation levels, which differ 
38: to large extent from those obtained using the replicator dynamics.
39: 
40: \end{abstract}
41: 
42: 
43: \maketitle
44: 
45: \section{Introduction}
46: \label{intro}
47: Understanding the emergence and persistence of cooperation is one of 
48: the central problems in evolutionary biology and socioeconomics  
49: \cite{MaynardSmith1995,Neumann1944}. In investigating this problem 
50: the standard framework utilized is evolutionary game theory  
51: \cite{Neumann1944,MaynardSmith1973,fudenberg}. Especially two models, the 
52: Prisoner's Dilemma \cite{Rapoport1965,Axelrod1981,Axelrod1988} and its  
53: variation, the snowdrift game \cite{MaynardSmith1973,Sugden1986}, 
54: have attracted most attention. In both games, the players can either 
55: cooperate for common good, or defect and exploit other players in 
56: attempt to gain benefits individually. In the Prisoner's Dilemma, 
57: the precondition is that it pays off  to be non-cooperative. 
58: Because of this, defection is the only evolutionarily stable 
59: strategy (ESS) in populations which are fully mixed, i.e. where each  
60: player interacts with any other player \cite{Smith1982}. However, 
61: several models which are extensions of the Prisoner's Dilemma have 
62: proved to sustain cooperation. These models include those in which 
63: the players are assumed to have memory of the previous interactions 
64: \cite{Axelrod1984}, or characteristics that allow cooperators and 
65: defectors to distinguish each other \cite{Epstein1998}, or players 
66: are spatially distributed \cite{hauert,Lindgren1997,Nowak1992}.  
67:  
68: A typical spatial game is such where player-player interactions only 
69: take place within restricted neighborhoods on regular lattices 
70: \cite{Nowak1992,Doebeli1998,Szabo1998,Szabo2000} or on complex 
71: networks \cite{Zimmermann2004}. These games have been found to 
72: generate highly complex behavior and enable the persistence of 
73: cooperation. Regarding the latter, the opposite was recently seen 
74: in the case of the snowdrift game played on a two-dimensional 
75: lattice \cite{hauert}, where the spatial structure resulted in 
76: decreased cooperator densities compared to the fully mixed  
77: ``mean-field'' case. This result was surprising, as intermediate 
78: levels of cooperation persist in unstructured snowdrift games, 
79: and the common belief has been that spatial structure is usually 
80: beneficial for sustained levels of cooperation. 
81:  
82: In these studies the viewpoint has largely been that of biological 
83: evolution, as represented by the so-called \emph{replicator dynamics}  
84: \cite{fudenberg,Hofbauer1998,Nowak2004}, where the fraction of players 
85: who use high-payoff-strategies grow (stochastically) in the population 
86: proportionally to the payoffs. This mechanism can be viewed as 
87: depicting Darwinian evolution, where the fittest have the largest 
88: chance of survival and reproduction. Overall, the factors influencing 
89: the outcomes of these spatially structured games are (i) the rules 
90: determining the payoffs (e.g. Ref.~\cite{fort2003}), (ii) the topology 
91: of the spatial structure (e.g. Ref.~\cite{Szabo2000}), and  (iii) the 
92: rules determining the evolution of each player's strategy (e.g. 
93: Ref.~\cite{Meyer1999,Traulsen2004}). We have studied the effect of 
94: changing  the strategy evolution rules (iii) in the two-dimensional 
95: snowdrift game similar  to that discussed in Ref.~\cite{hauert}. 
96: In our version, the rules have been defined in such a way that changes 
97: in the players' strategies represent player \emph{decisions} instead 
98: of different strategy genotypes in the next evolutionary generation 
99: of players. Thus, the time scale of the population dynamics in our  
100: model can be viewed to be much shorter than evolutionary time scales.  
101: Instead of utilizing the evolution-inspired replicator dynamics, 
102: we have endowed the players with primitive ``intelligence'' in the 
103: form of local decision-making rules determining their strategies. 
104: We show with simulations and analytic approach that these rules 
105: result in cooperation levels which differ largely from those 
106: obtained using the replicator dynamics.  
107: 
108: In this study we will concentrate on an adaptive snowdrift game, 
109: with agents interacting with their nearest neighbor agents on a 
110: two-dimensional square lattice. In what follows we first describe 
111: our spatial snowdrift model and then analyze its equilibrium states. 
112: Next we present our simulation results and finally draw some 
113: conclusions. 
114:  
115: \section{Spatial Snowdrift Model} 
116: \label{sec:2}  
117: The snowdrift model\footnote{Commonly known as hawk-dove or chicken
118:   game also.} can be illustrated with a situation in which 
119: two cars are caught in a blizzard and there is a snowdrift blocking 
120: their way. The cars are equipped with shovels, and the drivers have  
121: two choices: either start shoveling the road open or remain in the 
122: car. If the road is cleared, both drivers gain the benefit $b$ of 
123: getting home. On the other hand, clearing the road requires some work, 
124: and cost $c$ can be assigned to it ($b>c>0$). If both drivers are 
125: cooperative and willing to shovel, this workload is shared between 
126: them, and both of them gain total benefit of $R=c-b/2$. If both choose 
127: to defect, i.e. remain in their cars, neither one gets home and 
128: thus both obtain zero benefit $P=0$. If only one of the drivers shovels, 
129: both get home, but the defector avoids the cost and gains benefit $T=b$, 
130: whereas the cooperator's benefit is reduced by the workload, i.e. $S=b-c$.  
131:  
132: The above described situation can be presented with the bi-matrix 
133: \cite{gibbons} (Table \ref{payoffs}), where   
134: \begin{equation} 
135: T > R > S > P. 
136: \label{rank} 
137: \end{equation}
138: In case of the so called one-shot game, each player has two available 
139: strategies, namely defect (D) or cooperate (C). The players choose their 
140: strategies simultaneously, and their individual payoffs are given by 
141: the appropriate cell of the bi-matrix. By convention, the payoff  
142: to the so-called row player is the first payoff given, followed by the payoff  
143: of the column player. Thus, if for example player 1 chooses D and player 2 
144: chooses C, then player 1 receives the payoff T and player 2 the payoff S. 
145:  
146: The best action depends on the action of the co-player such that defect if 
147: the other player cooperates and cooperate if the other defects. A simple 
148: analysis shows that the game does not have \emph{stable evolutionary strategy} 
149: \cite{Hofbauer1998}, if the agents use only pure strategies, i.e., 
150: they can choose either to cooperate or to defect with probability 
151: one, but they are not allowed to use a strategy which mixes either 
152: of these actions with some probability $q \in (0,1)$. This leads to 
153: stable existence of cooperators and defectors in well-mixed 
154: populations \cite{hauert}. 
155: 
156: \begin{table} 
157: \caption{Snowdrift game. Player 1 chooses an action from the rows and 
158:   player 2 from the columns. By convention, the payoff to the row 
159:   player is the first payoff given, followed by the payoff of the column 
160:   player.}
161: \label{payoffs} 
162: \begin{center} 
163: \begin{tabular}{lll}
164: \hline\noalign{\smallskip} 
165:   & D   & C     \\ 
166: \noalign{\smallskip}\hline\noalign{\smallskip}
167: D & P, P & T, S \\ 
168: C & S, T & R, R \\ 
169: \noalign{\smallskip}\hline
170: \end{tabular} 
171: %\vspace*{5cm}
172: \end{center} 
173: \end{table} 
174:  
175: In order to study the effect of spatial structure on the snowdrift game,  
176: we set the players on a regular two-dimensional square lattice 
177: consisting of $m$ cells. We adopt the notation of Ref.~(\cite{schweitzer}) 
178: and identify each cell by an index $i=1,\ldots,m$ 
179: which also refers to its spatial position. Each cell, representing a player, 
180: is characterized by its strategy $s_i$, which can be either to cooperate  
181: ($s_i=1$) or to defect ($s_i=0$). The spatio-temporal distribution of the  
182: players is then described by $S=(s_1,\ldots,s_m)$ which is an element of 
183: a $2^m$ dimensional hypercube. Then every player -- henceforth called an 
184: \emph{agent} -- interacts with their $n$ nearest neighbors. We use either 
185: the Moore neighborhood in which case each agent has $n=8$ neighbors, in 
186: N,NE,E,SE,S,SW,W and NW, or the von Neumann neighborhood in which case 
187: each agent has $n=4$ neighbors, in N,E,S and W compass directions \cite{adachi}. 
188: We require that an agent plays \emph{simultaneously} with all its $n$ 
189: neighbors, and define the payoffs for this $(n+1)-player$ game  
190: such that an agent $i$ who interacts with $n_c^i$ 
191: cooperators and $n_d^i$ defectors, $n_c^i+n_d^i=n$, gains a benefit of 
192: \begin{eqnarray} 
193: u_i(s_i=0) & = & n_c^iT + n_d^iP \label{neiguta}\\ 
194: u_i(s_i=1) & = & n_c^iR + n_d^iS,  
195: \label{neigutb} 
196: \end{eqnarray}  
197: from defecting or cooperating, respectively.  
198:  
199: For determining their strategies, the agents are endowed with primitive  
200: decision-making capabilities. The agents retain no memory of the past, and  
201: are not able to predict how the strategies of the neighboring agents will  
202: change. Every agent simply assumes that the strategies of other agents  
203: within its neighborhood remain fixed, and chooses an action that maximizes  
204: its own payoff. In this sense the agents are myopic. The payoff is maximized, if an agent (a) defects when  
205: $u_i(0) > u_i(1)$, and (b) cooperates when $u_i(1) > u_i(0)$. If (c)  
206: $u_i(0) = u_i(1)$ the situation is indifferent. Using
207: Eqs.~(\ref{neiguta}) and (\ref{neigutb})  
208: we can connect the preferable choice of an agent and the payoffs of the game.  
209: Let us denote  
210: \begin{equation} 
211: \frac{1}{r}=1+\frac{S-P}{T-R}. 
212: \label{cbr} 
213: \end{equation} 
214: Then, if 
215: \begin{eqnarray} 
216: \frac{n_c^i}{n} & > & 1-r \textrm{ defecting is 
217:   profitable, or if} \label{boundarya}\\ 
218: \frac{n_c^i}{n} & < & 1-r \label{boundaryb}\textrm{ cooperating is 
219:   profitable, or if} \\ 
220: \frac{n_c^i}{n} & = & 1-r \textrm{ choices are indifferent}. 
221: \label{boundaryc} 
222: \end{eqnarray}  
223:  
224: Thus, for each individual agent, the ratio $r$ determines a following 
225: decision-boundary 
226: \begin{equation} 
227: \theta = n(1-r), \label{theta} 
228: \end{equation} 
229: which depends on the neighborhood size $n$ and the ``temptation''  
230: parameter $r$. Because $r$ is determined only by the differences  
231: $T-R$ and $S-P$, we can fix two of the payoff values, say $R=1$ and $P=0$.  
232: Based on the above, we define the following rules for the agents: 
233: \begin{enumerate} 
234: \item If an agent $i$ plays at time $t$ a strategy $s_i(t) \in \{0,1\}$ for 
235:   which $u_i(s_i) \geq u_i(1-s_i)$, then at time $t+1$ the agent plays 
236: $s_i(t+1) =  s_i(t)$. 
237: \item  If an agent $i$ plays at time $t$ a strategy $s_i(t) \in \{0,1\}$ for 
238:   which $u_i(s_i) < u_i(1-s_i)$, then at time $t+1$ the agent plays  
239: $s_i(t+1) = 1  - s_i(t)$ with probability $p$, and $s_i(t+1)=s_i(t)$  
240: with probability $1-p$. 
241: \end{enumerate} 
242: Hence, the strategy evolution of an individual agent is determined by  
243: the current strategies of the other agents within its neighborhood, with 
244: the parameter $p$ acting as a ``regulator'' which moderates the rate 
245: of changes.  
246:  
247:  
248: \section{Equilibrium states} 
249: \label{sec:3}
250: A spatial game is in stable state or equilibrium if retaining the current 
251: strategy is beneficial for all the agents~\cite{fudenberg}. There can be  
252: numerous equilibrium configurations, depending on the temptation  
253: parameter $r$, geometry and size of the $n$-neighborhood, 
254: and the size and boundary conditions of the lattice upon which the 
255: game is played. An aggregate quantity of particular interest is the 
256: fraction of cooperators $F_c$ in the whole population (or, equivalently, 
257: that of the defectors $F_d$). Below, we derive limits for $F_c$, first 
258: in a ``mean-field'' picture based cooperator densities within neighborhoods 
259: and then by investigating local neighborhood configurations. 
260:  
261: \subsection{Mean-field limits for cooperator density} 
262: \label{sec:3a} 
263: Without detailed knowledge of local equilibrium configurations we can 
264: already derive some limits for the fraction of cooperators in equilibrium. 
265: Let us consider a square lattice with $m=L\times L$ cells with periodic 
266: boundary conditions, where $L$ is the linear size of the lattice, and 
267: assume that $k$ cells are occupied by cooperators. We denote by  
268: $a_j$ the number of those agents who have $j$ cooperators each in their 
269: $n$-neighborhood, excluding the agents themselves, and denote the local 
270: density of cooperators in such neighborhoods by $f_c=j/n$. Hence, the 
271: total amount of cooperators $k$ can be written in terms of the 
272: densities as follows 
273:  
274: \begin{equation} 
275: k = \sum_{j=0}^n a_j f_c = \sum_{j=0}^n a_j \frac{j}{n}. 
276: \label{id1} 
277: \end{equation} 
278:  
279: From Eqs.~(\ref{boundarya})-(\ref{boundaryc}) we can infer that a cooperator will retain 
280: its current strategy, if it has at most $c$ cooperators in its 
281: $n$-neighborhood, where $c$ is the integer part of $\theta=n(1-r)$. 
282: Similarly, a defector will remain a defector if it has more than $c$ 
283: cooperators in its neighborhood. Thus, in equilibrium, all agents 
284: having $j\leq c$ cooperators in their neighborhood are likewise 
285: cooperators, and thus $\sum_{j=0}^c a_j = k$. We denote by 
286: $\left<f_{c|c}\right> = \frac{1}{k}\sum_{j=0}^c a_j \frac{j}{n}$ 
287: the average density of cooperators as the nearest neighbors of 
288: cooperators. Similarly, $\left<f_{c|d}\right>$ denotes the average 
289: density of cooperators as the nearest neighbors of defectors, i.e. 
290: $\left<f_{c|d}\right> = \frac{1}{m-k}\sum_{j=c+1}^n a_j \frac{j}{n}$.  
291: Then we can write Eq.~(\ref{id1}) as 
292: \begin{equation} 
293: k=k\left<f_{c|c}\right> + (m-k)\left<f_{c|d}\right>. 
294: \label{id2} 
295: \end{equation} 
296:  
297: \begin{figure} 
298: \begin{center}
299: \resizebox{0.4\textwidth}{!}{
300: \includegraphics{figure1.eps}} 
301: \caption{In equilibrium the average density of cooperators in the 
302:   nearest neighborhood of defectors must be $1-r \leq \left < f_{c|d} \right > \leq 1$ and 
303:   in the nearest neighborhood of cooperators $0 \leq \left < f_{c|c} \right > \leq 1-r$ 
304:   (shaded area). If the total number of players in the lattice is $m$, 
305:   the lines $k\left < f_{c|c} \right > + (m-k)\left < f_{c|d} \right > 
306:   = k$ depict the identity of $k$ cooperators in the lattice. Equilibrium is 
307:   not possible, if the fraction of cooperators $F_c = k/m$ is such that the lines 
308:   do not pass through the shaded area.\label{fig1}} 
309: \end{center} 
310: \end{figure} 
311:  
312: The density  $f_{c|c}$ of cooperators around each cooperator is bounded: 
313: $f_{c|c}\geq 0$, $f_{c|c}\leq c/n$, and as $c \leq \theta = n(1-r)$, the 
314: relation $0 \leq \left<f_{c|c}\right> \leq 1-r$ holds for the average density. 
315: Similarly, the density of cooperators around each defector $f_{c|d}$ can be 
316: at most $1$ and is at least $(1-r)$, and thus the average density 
317: $1-r \leq \left<f_{c|d}\right> \leq 1$. Using these relations together 
318: with Eq.~(\ref{id2}) we obtain the following limits for the density of 
319: cooperators $F_c=k/m$ in the whole agent population 
320: (see also Fig.~\ref{fig1}): 
321: \begin{equation} 
322: \frac{1-r}{2-r} \leq F_c \leq \frac{1}{r+1}. 
323: \label{loup} 
324: \end{equation} 
325:  
326: \subsection{Local equilibrium configurations} 
327:  \label{sec:3b}
328: In the above derivation we ignore how the strategies can actually be 
329: distributed in the lattice. Hence, it is of interest to examine possible 
330: local equilibrium configurations of the player strategies. Again, 
331: Eqs.~(\ref{boundarya})-(\ref{boundaryc}) tell us how many cooperative neighbors each 
332: defector or cooperator can have in the equilibrium state. The number 
333: of cooperators around each agent depends on the value of the temptation 
334: parameter $r$, and for a given value of $r$ the lattice has to be filled 
335: such that these conditions hold for the neighborhood of each agent. 
336: In a lattice with periodic boundary conditions, the lattice size 
337: $m=L_X \times L_Y$ and the neighborhood size $n$ obviously have an 
338: effect on the elementary configurations. Hence, we restrict ourselves 
339: to infinite-sized lattices, filled by repeating elementary configuration 
340: blocks, and look for the resulting limits on the cooperator density $F_c$.  
341: Note that these conclusions also hold for finite lattices with periodic 
342: boundary conditions, if $L_X$ and $L_Y$ are integer multiples of $X$ 
343: and $Y$, respectively, where $X\times Y$ is the elementary block size. 
344: Here, we will restrict the analysis to the case of the Moore neighborhood 
345: with $n=8$. 
346:  
347: \begin{table} 
348: \caption{ Limits for the equilibrium fraction of cooperators based on 
349:         repeating elementary configuration blocks. When $r_l < r < r_u$, 
350:   the number of cooperators in each defector's neighborhood $N_{c|d}$  
351:   must be at least $9-i$ and the number of cooperators in each 
352:   cooperator's neighborhood $N_{c|c}$ at most $8-i$. Considering 
353:   possible repeating configuration blocks which fulfill these 
354:   conditions, we obtain lower limits $F_{c,L}$ and upper limits 
355:   $F_{c,U}$ for the density of cooperators.} 
356: \label{tab1} 
357: \begin{center} 
358: \begin{tabular}{lllccll}
359:  \hline\noalign{\smallskip}
360: i &$r_l$  & $r_u$ & $N_{c|d} \ge$ & $N_{c|c} \le$ & $F_{c,L}$ &
361: $F_{c,U}$ \\ 
362: \noalign{\smallskip}\hline\noalign{\smallskip} 
363: 1&$0$   & $1/8$ & 8  & 7  & $3/4$  & $8/9$ \\ 
364: 2&$1/8$ & $2/8$ & 7  & 6  & $2/3$  & $4/5$ \\ 
365: 3&$2/8$ & $3/8$ & 6  & 5  & $1/2$  & $2/3$ \\ 
366: 4&$3/8$ & $4/8$ & 5  & 4  & $1/2$  & $2/3$ \\ 
367: 5&$4/8$ & $5/8$ & 4  & 3  & $4/9$  & $1/2$ \\ 
368: 6&$5/8$ & $6/8$ & 3  & 2  & $1/3$  & $1/2$ \\ 
369: 7&$6/8$ & $7/8$ & 2  & 1  & $2/9$  & $1/3$ \\ 
370: 8&$7/8$ & $8/8$ & 1  & 0  & $1/9$  & $1/4$ \\ 
371: \noalign{\smallskip}\hline
372: \end{tabular} 
373: \end{center} 
374: \end{table} 
375:  
376: \begin{figure} 
377: \begin{center} 
378: \resizebox{0.4\textwidth}{!}{
379: \includegraphics{figure2.eps}} 
380: \caption{ Examples of elementary configuration blocks which can be 
381:   repeated without overlap to fill an infinite lattice, for various  
382:   values of $r$. The numbering refers to $i$ in Table \ref{tab1}. A 
383: black cell denotes a defector while an empty cell denotes a cooperator. 
384: For a particular number the lower limit of density is obtained by filling 
385: the lattice with the blocks on the left, and the upper by using the 
386: blocks on the right. \label{fig2}} 
387: \end{center} 
388: \end{figure} 
389:  
390: As an example, consider the local configurations when $r=0.1$,  
391: and hence the decision boundary value $\theta = n(1-r) = 7.2$.  
392: Thus, from Eqs.~(\ref{boundarya})-(\ref{boundaryc}) one can infer that in equilibrium 
393: all defectors should have more than $7.2$ cooperators in their 
394: Moore neighborhoods. Because the number of cooperating neighbors 
395: can take only integer values, this means that every one of the $n=8$ 
396: neighbors of a defector should be a cooperator. On the other hand, 
397: from Eqs.~(\ref{boundarya})-(\ref{boundaryc}) we see that the density $f_{c|c}$ of 
398: cooperators around each cooperator should be less than $1-r$, i.e. 
399: they should have at most $c=7$ cooperators in their Moore neighborhood. 
400: The smallest repeated elementary block fulfilling both conditions 
401: is a $2\times 2$-square with one defector -- when the lattice  
402: is filled with these blocks, the cooperator density equals $F_c=3/4$ 
403: (see Fig.~(\ref{fig2}), case 1, left block). On the other hand, both 
404: requirements are likewise fulfilled with a repeated $3\times 3$-square, 
405: where the central cell is a defector and the rest are cooperators, 
406: resulting in the cooperator density of $F_c=8/9$. This configuration 
407: is illustrated in Fig.~(\ref{fig2}), as case 1, right block. 
408:    
409: By continuing the analysis of elementary configuration blocks in similar 
410: fashion for different values of $r$, we obtain lower and upper limits 
411: for the fraction of cooperators, which are listed in Table \ref{tab1}.  
412: The corresponding elementary configuration blocks are depicted 
413: in Fig.~(\ref{fig2}). The table is read so that when the value of 
414: the temptation parameter is within the interval $r_l < r < r_u$,  
415: the number of cooperators in each defector's neighborhood $N_{c|d}$ 
416: must be at least $9-i$ and the number of cooperators in each 
417: cooperator's neighborhood $N_{c|c}$ can be at most $8-i$. Here
418: $r_l=(i-1)/8$, $r_u=i/8$ and $i=1,\ldots,8$ These conditions are 
419: those of Eqs.~(\ref{boundarya})-(\ref{boundaryc}) and they are fulfilled by the 
420: configuration blocks depicted in Fig.~(\ref{fig2}), for which the 
421: minimum and maximum densities of cooperators are $F_{c,L}$ and $F_{c,U}$. 
422:  
423: \section{Simulation results} 
424:  \label{sec:4}
425: We have studied the above described spatial snowdrift model with 
426: discrete time-step simulations on a $m=100 \times 100$-lattice with 
427: periodic boundary conditions. We have specifically analyzed the behavior 
428: of the cooperator density $F_c$, and equilibrium lattice configurations. 
429: In the simulations, the lattice is initialized randomly so that each  
430: cell contains a cooperator or defector with equal probability. However, 
431: biasing the initial densities toward cooperators or defectors was found to  
432: have no considerable effect on the outcome of the game. We have simulated 
433: the game using both the Moore and the von Neumann neighborhoods 
434: with $n=8$ and $n=4$ nearest neighbors, respectively. In the simulations 
435: we update strategies of the agents asynchronously \cite{adachi} with the 
436: random sequential update scheme, so that during one simulation round, 
437: every agent's strategies are updated in random order. In the following, 
438: the time scale is defined in terms of these simulation rounds. 
439:  
440: First, we have studied the development of the cooperator density  
441: $F_c$ as a function of time. As expected, the probability 
442: $p$ of discontent agents changing their strategies plays the role  
443: of defining the convergence time scale only\footnote{The 
444: role of $p$ would be more important if synchronous update rules were 
445: used. In that case $p=1$ corresponds to a situation where each 
446: discontent agent simultaneously changes its strategy to  
447: the opposite. This, then, could result in a frustrated situation 
448: with oscillating cooperator density. However, 
449: small enough values of $p$ should damp these oscillations, 
450: resulting in static equilibrium.}, as in the long run $F_c$ converges 
451: to a stable value irrespective of $p$. This is depicted in Fig.~\ref{fig3}, 
452: which shows $F_c$ as function of time for several values of $p$ and 
453: two different values of the temptation $r$. In these runs, we have used 
454: the Moore neighborhood, i.e. $n=8$. In all the studied cases, $F_c$ 
455: turns out to converge quite rapidly to a constant value, $F_c\sim 0.7$ 
456: for $r=0.2$ and $F_c\sim 0.3$ for $r=0.8$. 
457:  
458: It should be noted that $F_c$ does not have to converge to  
459: exactly the same stable value for the same $r$; even if the game 
460: is considered to be in equilibrium, there can be some variance 
461: in $F_c$, which is also visible in Fig.~\ref{fig3}. However, 
462: the value of $F_c$ was found to eventually remain stable during 
463: individual runs, i.e. no oscillations were detected.  
464:  
465: %Fig.~(\ref{fig3}) also points out that if we use the Moore 
466: %neighbourhood and pick $r$ from two different intervals 
467: %$i/8<r<(i+1)/8$, i.e., for two different $i = 1,\ldots,7$, then the values of 
468: %$f_c$ tend to converge to different levels. Partly, the explanation for this 
469: %is given by Eqs.~(\ref{boundarya})-(\ref{boundaryc}) which show how the number of 
470: %cooperators in each agent's neighbourhood depends on 
471: %$r$. We discussed about this dependency already in Equilibrium section 
472: %when we derived upper and lower limits for $f_c$. We can also 
473: %check that $f_c$ in Fig.~(\ref{fig3}) falls between the lower and 
474: %upper limits given in Tab.~(\ref{tab1}). For $r=0.2$, 
475: %$i=2$ in the above interval. From Tab.~(\ref{tab1}) we see that 
476: %corresponding to $i=2$ the lower limit $f_l(c)=2/3$ and the upper 
477: %limit $f_l(c)=4/5$. In Fig.~(\ref{fig3}) $f_c$ converges to values 
478: %slightly over $0.7$ when $r=0.2$ for all used $p$. This clearly falls 
479: %between the lower and upper limits. Similarly we can check that for 
480: %$r=0.8$, $i=7$ in Tab.~(\ref{tab1}), and the lower and upper limits for 
481: %$f_c$ are $f_l(c)=2/9$ and $f_u(c)=1/3$. In Fig.~(\ref{fig3}) the 
482: %curves for $r=0.8$ fall between these limits. 
483:  
484: \begin{figure} 
485: \begin{center}
486: \resizebox{0.4\textwidth}{!}{
487: \includegraphics{figure3.eps}} 
488: \caption{Dynamics of the fraction of cooperators $F_c$. The upper curves  
489: that converge to $F_c \sim 0.7$ are for $r=0.2$, and the lower curves
490: that converge to $F_c \sim 0.3$ are for $r=0.8$. In both cases the probability of 
491: being discontent is varied as $p=1,0.1,0.01,0.001$ 
492: from left to right, and the lattice size is $m=100x100$.} 
493: \label{fig3} 
494: \end{center} 
495: \end{figure} 
496:  
497: Next, we have studied the average equilibrium fraction of cooperators 
498: $\left< F_c\right >$ in the agent population as function of the temptation 
499: parameter $r$. We let the simulations run for 500 rounds (with $p=0.1$), and 
500: averaged the fraction of cooperators for the subsequent 500 rounds. 
501: In all cases, the fraction had already converged before the averaging rounds. 
502: Fig.~(\ref{fig4}) shows the results for the von Neumann neighborhood ($n=4$), illustrated as
503: the squares. The dotted lines indicate the upper and lower limits of  
504: Eq.~(\ref{loup}), and the dashed diagonal line is $F_c=1-r$, corresponding 
505: to the fraction of cooperators in the fully mixed 
506: case~\cite{fudenberg,hauert,Hofbauer1998}. The fraction of cooperators 
507: $\left<F_c\right>$ is seen to follow a stepped curve, with steps corresponding 
508: to $r=i/n$, where $i=0,\ldots,n$. This is a natural consequence of 
509: Eqs.~(\ref{boundarya})-(\ref{boundaryc}), where the decision boundary $\theta=n(1-r)$ 
510: can take only discrete values. A similar picture is given for the Moore neighborhood 
511: ($n=8$) in the middle panel of Fig.~(\ref{fig5}). Furthermore, in the
512: middle panel of Fig.~(\ref{fig5}) the values 
513: of $F_c$ fall between the limits given in Table \ref{tab1} for all $r$
514: as shown with solid lines. 
515:   
516: In both cases (i.e. with Moore and von Neumann neighborhoods) cooperation is 
517: seen to persist during the whole range $r=[0,1]$. This result differs 
518: largely from the $F_c(r)$-curves of the spatial snowdrift game with 
519: replicator dynamics~\cite{hauert}, where the fraction of cooperators 
520: vanished at some critical $r_c$. Hence, we argue that no conclusions 
521: on the effect of spatiality on the snowdrift game can be drawn without 
522: taking into consideration the strategy evolution mechanism; local 
523: decision-making in a restricted neighborhood yields results which are 
524: different from those resulting from the evolutionary replicator dynamics.   
525:  
526: \begin{figure} 
527: \begin{center}
528: \resizebox{0.4\textwidth}{!}{
529: \includegraphics[angle=270]{figure4.eps}}
530: \caption{Average fraction of cooperators $\left<F_c\right>$ versus the
531:   temptation $r$ (squares), 
532: simulated on a $100\times 100$ lattice with $p=0.1$ using the von
533: Neumann neighborhood. The values for $\left<F_c\right>$ 
534: are averages over 500 simulation rounds, where the averaging was started after 500 
535: initial rounds to guarantee convergence. The dotted lines depict the upper and 
536: lower limits for $F_c$ of Eq.~(\ref{loup}). The dashed diagonal line is $1-r$.} 
537: \label{fig4} 
538: \end{center} 
539: \end{figure} 
540:  
541: We have also studied the equilibrium lattice configurations for  
542: various values of $r$. Fig.~(\ref{fig5}) depicts the central part of the  
543: $100\times 100$-lattice after 1000 simulation rounds using the Moore  
544: neighborhood and $p=0.1$, with white pixels corresponding to cooperators and 
545: black pixels to defectors. The values of $r$ have been selected so that 
546: the equilibrium situation corresponds to each plateau of
547: $\left<F_c\right>$ illustrated in the central panel.
548:  
549: The observed configurations are rather polymorphic, and repeating elementary 
550: patterns like those in Fig.~(\ref{fig2}) are not seen. This reflects the fact 
551: that the local equilibrium conditions can be satisfied by various configurations; 
552: the random initial configuration and the asynchronous update then lead to 
553: irregular-looking equilibrium patterns, which vary between simulation runs. 
554: The patterns seem to be most irregular when $r$ is around 0.5; this is because 
555: then the equilibrium numbers of cooperators and defectors are close to each other, 
556: and the ways to assign strategies within local neighborhoods are most numerous. 
557: To be more exact, there are $8 \choose i$ ways to distribute $i$ cooperators in 
558: the $8$-neighborhood, and if e.g. $3/8<r<4/8$, $i$ is at least $4$ and at most $5$, 
559: maximizing the value of the binomial coefficient. Hence, the ways of filling the 
560: lattice with these neighborhoods in such a way that the equilibrium conditions 
561: are satisfied everywhere are most numerous as well. 
562:  
563:  
564: \begin{figure*} 
565: \begin{center}
566: \resizebox{0.75\textwidth}{!}{
567: \includegraphics{figure5.eps}}
568: \caption{ Example equilibrium configurations of defectors and cooperators on a 
569:  $m=100\times 100$ lattice for various values of $r$ when the Moore
570:  neighborhood is used. The configurations 
571: were recorded after $T=1000$ simulation rounds. Only the middle part of  
572: the lattice is shown for the sake of clarity. The middle panel depicts the 
573: average fraction of cooperators $\left<F_c\right>$ in the whole population as a function 
574: of the temptation $r$ (squares), together with the upper and lower limits of 
575: Eq.~(\ref{loup}) (dotted lines) and the limits of Table \ref{tab1} (solid
576: lines). The values of $\left<F_c\right>$ are averages over the last
577: 500 simulation rounds and the dashed diagonal line is $F_c=1-r$, corresponding 
578: to the fraction of cooperators in the fully mixed case~\cite{fudenberg,hauert,Hofbauer1998}.} 
579: \label{fig5} 
580: \end{center} 
581: \end{figure*} 
582:  
583:  
584: \section{Summary and conclusions} 
585:  \label{sec:5}
586: We have presented a variant of the two-dimensional snowdrift game, where the 
587: strategy evolution is determined by agent decisions based on the strategies 
588: of other players within its local neighborhood. We have analyzed the lower 
589: and upper bounds for equilibrium cooperator densities with a mean-field 
590: approach as well as considering possible lattice-filling elementary configuration 
591: blocks. We have also shown with simulations that this game converges to 
592: equilibrium configurations with constant cooperator density depending on 
593: the payoff parameters, and that these densities fall within the derived limits. 
594: Furthermore, the strategy configurations in the equilibrium state display 
595: interesting patterns, especially for intermediate temptation parameter values. 
596:  
597: Most interestingly, the equilibrium cooperator densities differ largely from those 
598: resulting from applying the replicator dynamics~\cite{hauert}. With our strategy 
599: evolution rules, cooperation persists through the whole temptation parameter 
600: range. This illustrates that one cannot draw general conclusions on 
601: the effect of spatiality on the snowdrift game without taking the strategy 
602: evolution mechanisms into consideration -- this should, in principle, apply 
603: for other spatial games as well. Care should especially be taken when 
604: interpreting the results of investigations on such games: the utilized 
605: strategy evolution mechanism should reflect the system under study. We argue 
606: that especially when modeling social or economic systems, there is 
607: no \emph{a priori} reason to assume that generalized conclusions can be 
608: drawn based on results using the evolution inspired replicator dynamics 
609: approach, where high-payoff strategies get copied and ``breed'' in proportion 
610: to their fitness. As we have shown here, local decision-making with limited 
611: information (neighbor strategies are known payoffs are not) can result 
612: in different outcome.
613: 
614: % Non-BibTeX users please use
615: \begin{thebibliography}{}
616: 
617: \bibitem{MaynardSmith1995}J. {Maynard Smith} and E. Szathm\'{a}ry,
618:   \textit{The Major Transitions in Evolution} (W.H. Freeman, Oxford, UK, 1995). 
619: 
620: \bibitem{Neumann1944}J. {von Neumann} and O. Morgenstern,
621:   \textit{Theory of Games and Economic Behaviour} (Princeton University Press, 1944).
622: 
623: \bibitem{MaynardSmith1973}J. {Maynard Smith} and G. Price, Nature
624:   \textbf{246} (1973) 15-18.
625: 
626: \bibitem{fudenberg}D. Fudenberg and D. K. Levine, \textit{The Theory
627:     of Learning in Games} (The MIT Press, 1998).
628: 
629: \bibitem{Rapoport1965}A. Rapoport and A. Chammah, \textit{Prisoner's
630:     Dilemma} (University of Michigan Press, Ann Arbor, 1965).
631: 
632: \bibitem{Axelrod1981}R. Axelrod and W.D. Hamilton, Science
633:   \textbf{211}, (1981) 1390-1396.
634: 
635: \bibitem{Axelrod1988}R. Axelrod and D. Dion, Science \textbf{242}, (1988) 1385-1390.
636: 
637: \bibitem{Sugden1986}R. Sugden, \textit{The Economics of Rights,
638:     Co-operation and Welfare} (Blackwell, Oxford, UK, 1986).
639: 
640: \bibitem{Smith1982}J.M. Smith, \textit{Evolution and the theory of
641:     games} (Cambridge University Press, Cambridge, UK, 1982).
642: 
643: \bibitem{Axelrod1984}R. Axelrod, \textit{The evolution of
644:     cooperation}, (Basic Books, New Yourk, 1984).
645: 
646: \bibitem{Epstein1998}J. N. Epstein, Complexity \textbf{4}(2), (1998) 36-48.
647: 
648: \bibitem{hauert}C. Hauert and M. Doebell, Nature \textbf{428}, (2004) 643-646.
649: 
650: \bibitem{Lindgren1997}K. Lindgren, \textit{Evolutionary Dynamics in
651:     Game-Theoretic Models in The Economy as an Evolving Complex System
652:     II} (Addison-Wesley, 1997) 337-367.
653: 
654: \bibitem{Nowak1992}M. A. Nowak and R. May, Nature \textbf{359} (1992) 826-829.
655: 
656: \bibitem{Doebeli1998}M. Doebeli and N. Knowlton, Proc. Natl Acad. Sci
657:   USA \textbf{95} (1998) 8676-8680.
658: 
659: \bibitem{Szabo1998}G. Szab\'{o} and C. Toke, Phys. Rev. E
660:   \textbf{58} (1998) 69-73.
661: 
662: \bibitem{Szabo2000}G. Szab\'{o}, T. Antal, P. Szab\'{o}  and M. Droz, Phys. Rev. E \textbf{62} (2000) 1095-1103. 
663: 
664: \bibitem{Zimmermann2004}M. G. Zimmermann, V. M. Egu\'{i}luz and M. San
665:   Miguel, Phys. Rev. E \textbf{69} (2004) 065102.
666: 
667: \bibitem{Hofbauer1998}J. Hofbauer and K. Sigmund, \textit{Evolutionary
668:     games and population dynamics}, (Cambridge University Press,
669:   Cambridge, UK, 1998). 
670: 
671: \bibitem{Nowak2004}M. A. Nowak and K. Sigmund, Science \textbf{303}
672:   (2004) 793-799.
673: 
674: \bibitem{fort2003}H. Fort, Phys. Rev. E (2003) \textbf{68} 026118.
675: 
676: \bibitem{Meyer1999}D. A. Meyer, Phys.Rev.Lett. \textbf{82} (1999) 1052-1055.
677: 
678: \bibitem{Traulsen2004}A. Traulsen, T. R\"{o}hl and H.G.Schuster,
679:   Phys.Rev.Lett. \textbf{93} (2004) 028701.
680: 
681: \bibitem{gibbons}R. Gibbons, \textit{Game Theory for Applied
682:     Economists}, (Princeton University Press, 1992). 
683: 
684: \bibitem{schweitzer}F. Schweitzer, L. Behera and H. M\"{u}hlenbein, Advances in Complex Systems \textbf{5} (2002) 269-299.
685: 
686: \bibitem{adachi}S. Adachi, F. Peper and J. Lee, Journal of Statistical
687:   Physics \textbf{114} (2004) 261-289.
688: 
689: \end{thebibliography}
690: 
691: 
692: \end{document}
693: 
694: