The Study On Autonomous Cooperation Of Muti-agent System With Memory Mechanism

ABSTRACT:It is a challenge to understand how cooperation can emerge in rational population in self-organized systems.Evolutional game theory provides a suitable theoretical framework to address this issue.In the research of cooperation in evolutionary Prisoner’s dilemma game, Snow drift game and Stag-hunt game,we proposed a novel systematic analysis chain of "cooperation ratio-fitness of the system-distribution of R, ST, P reciprocity-distribution of strategies".This analysis chain can help us investigating and understanding the process of the system’s evolution more clearly. The main contend and contributions of this dissertation are summarized as follows:Firstly, we investigated the memory’s effect to cooperation.We discussed the difficulty on mathematic way to analyze the system with memory in well-mixed population.Then by the multi-agent model,we find that memory has different effects to different dilemma games.Memory can promote cooperation greatly in Prisoner’s Dilemma game.Memory’s effect to Snowdrift game includes2cases:when the temptation of betray is small,memory demote cooperation; when the temptation of betray is great, memory can inhibit the temptation and promote cooperation.Memory can promote cooperation in Stag hunt game in most cases.Our researches also find that although cooperation is partly demoted in Snowdrift game,ST reciprocity is actually enhanced and R reciprocity is weakened. So the system’s fitness increases while the cooperation decreases.We also find that strategy WSLS (Win-stay-lose-shift) can be widely adopted in Prisoner’s dilemma, Snowdrift game and Stag hunt game.Secondly, we studied the effect of initial memory of every new generation to cooperation.What interesting is that the systems remembering history cannot perform better than those randomly initialize the history in every new generation.Furthermore, to the unconditional imitation update rule,remembering the history demotes cooperation in Prisoner’s dilemma, Snowdrift game and Stag hunt game on the contrary. The cooperation level of these systems is far lower than those randomly initialized the history at the beginning of every new generation. To the replicator rule, systems who remembering history are similar to those who randomly initialized with50%cooperators.Different initial cooperation ratio affect the final cooperation greatly with replicator update rule, in which high initial cooperation ratio can lead high final cooperation;while in systems with unconditional imitation rule,different initial cooperation ratio affect final cooperation slightly. The fundamental reason for these phenomenon is that random initialization destroys the structure of some defect strategies.Thirdly, we investigated the cooperative turn-taking (CAD, cooperation alternating defection) in ST reciprocity and point out that it is another kind of cooperation.We built an extended S-T plane, and changed the memory to different length. We find that when agents can memory the last2step history, cooperative turn taking emerges on the area that S+T>2R. At the same time,another case of ST reciprocity-consistent defection or cooperation (DorC) emerges on the line S=T. We also investigated different memory length’s effect to cooperation level, system fitness on the extended S-T plane.Last we did some statistical analysis to the6-step history, which can verify the distribution of cooperative turn taking CAD and consistent defection or cooperation (DorC).Lastly we investigated the memory’s effect to the system in which agents can move randomly. We found that when there is no memory, to the mobility mode COD (combat-offspring-diffuse),low speed movement can promote cooperation in Prisoner’s Dilemma, Snow drift game and Stag-hunt game,while high speed demote cooperation; to the mobility mode CDO (combat-diffuse-offspring),what interesting is, high speed promote cooperation on the contrary. Mobility promote ST reciprocity in Snow drift game.Memory promote cooperation in COD mobility mode;while in CDO mode,the effects are complicated.