Decentralized Interaction and Co-Adaptation in the Repeated Prisoner&2018;s Dilemma

A Prisoner&2018;s dilemma that is repeated indefinitely has many equilibria; the problem of selecting among these is often approached using evolutionary models. The background of this paper is a number of earlier studies in which a specific type of evolutionary model, a genetic algorithm (GA), was used to investigate which behavior survives under selective pressure. However, that normative instrument searches for equilibria that may never be attainable. Furthermore, it aims for optimization and, accordingly, says what people should do to be successful in repeated prisoner&2018;s dilemma (RPD) type situations. In the current paper, I employ simulation to find out what people would do, whether this makes them successful or not. Using a replication of Miller&2018;s (1988) GA study for comparison, a model is simulated in which the population is spatially distributed across a torus. The agents only interact with their neighbors and locally adapt their strategy to what they perceive to be successful behavior among those neighbors. Although centralized GA-evolution may lead to somewhat better performance, this goes at the cost of a large increase in required computations while a population with decentralized interactions and co-adaptation is almost as successful and, additionally, endogenously learns a more efficient scheme for adaptation. Finally, when the agents&2018; perceptive capabilities are limited even further, so that they can only perceive how their neighbors are doing against themselves, rather than against all those neighbors&2018; opponents&2014;which essentially removes reputation as a source of information&2014;cooperation breaks down.