Simulation of multi-agent Prisoners' Dilemmas

This article is a report on computer simulation experiments based on our new agent-based simulation tool to model social situations for the case of large numbers of not necessarily rational decision-makers. It is a simulation of the multiagent Prisoners' Dilemma game for realistic (nondyadic) interactions between agents with various personalities. Our model has a number of user-defined parameters such as the size and shape of the simulation environment, the definition of neighborhood, the payoff (reward/penalty) functions, the learning rules, the agents' personalities, and the initial conditions. We have performed a series of simulation experiments with various combinations of these parameters. For the case of one class of agents we found two distinctly different nontrivial but remarkably regular solutions. For a wide range of initial conditions, the number of cooperators oscillates around a relatively small value. When the initial aggregate cooperation probability is above a certain value, the solutions tend to reach well-defined constant values that are dependent on the initial values. For other types of agents the solutions show interesting chaos-like behavior. The article provides some insight into the conditions of decentralized cooperation in spatially distributed populations of agents.