Game Theory and Agents

A fundamental problem in multi agent systems is conflict resolution. A conflict occurs in general when the agents have to deal with conflicting goals, such as demands for shared, but limited resources. We investigate how game theory may be a helpful and efficient tool for examining a class of conflicts in multi agent systems and argue that parameterized games and conclusions based on general properties of strategies are of interest. The thesis includes four papers, the former three describing various parameterized simulations, and the latter one describing the theory behind it and stating some remarks given by our approach. The first paper describes a parameterized hawk-and-dove game simulation that spans from the Chicken Game (CG) to the Prisoner’s Dilemma (PD). We conclude that the CG to a higher extent rewards cooperative strategies than PD, and that the change in score is linear to the change of the payoff matrix. In the second paper, we introduce the notion of generous and greedy strategies as being the ones that, given a certain distribution of opponents, play more (C,D) than (D,C) and more (D,C) than (C,D) respectively ((X,Y) is the move in which the opponent acts Y, when the own player acts X). The result, when simulated in CG and PD is that strategies starting as generous will outperform the greedy ones in a CG with a population game, while the opposite happens in the PD. The last of the simulation based papers, compare four different PD like games and states how the generous and greedy strategies manage in the different games. The fourth paper discuss why iterated games are interesting for multi agent research and how it is possible to use a pragmatic point of view in order to make an agent decide which strategy (or behavioral pattern) to use. The notion of Characteristic Distributions (ChDs) are introduced and it is used to support the agents in their choice. Based on the ChDs a “No Free Lunch” theorem for game theory is formulated and proved.