Modèles et simulations informatiques des problèmes de coopération entre agents. (Models and computer simulations of agents cooperation problems)

The iterated prisoner’s dilemma is a cooperation between agents mathematical representation. This modelcomes from game theory, which initial goal is to study conflict of interest situations. Unfortunately resultsimplied by the theory on the prisoner’s dilemma are so poor that it could not be used seriously. A newevolutionary approach, mainly based on computer simulations, has been initiated by Robert Axelrod. Agentsare characterized by their behavior, or strategy. Axelrod brings to the fore four properties a strategy has torespect in order to be efficient, and propose the tit for tat strategy as an example.Our work is to study and increase knowledge on those kind of simulations. We adapt the model such thatit takes into account the fact that computing is a discrete job. This adaptation allows us to make a lot ofsimulations which confirm main results obtained in the continuous case. This, however, call into question oneof Axelrod property : simplicity. We describe strategies more efficient than tit for tat but with a higherlevel of complexity. Evaluations are done by simulations involving a big number of strategies build objectivelyvia a genetic approach.Those simulations allow us to find a new property : behavior adaptation faculty. This new property enforcethe idea that cooperative behavior between agents may be more and more complex.We also start the study of a particular prisoner’s dilemma, which differs from the classical model only inthe iterated version and which allows to represent two level of cooperation : the lift dilemma. This theoreticalas well as experimental study allow us to show that with this new model no purely deterministic behavior canbe efficient.