Decentralized Interaction and Co-adaptation in the Repeated Prisoner's Dilemma

Tomas Klos - University of Groningen

This paper investigates the behavioral patterns that emerge from the interactions among boundedly rational adaptive agents. They interact in repeated prisoner's dilemma's (RPD's) and adapt their behavior after each RPD tournament. Results are compared across virtual experiments with different regimes of interaction and adaptation. Specifically, round-robin interactions and centralized evolution of the population is compared to locally interacting and co-adaptating agents on a torus. Furthermore, the effects of imposing an additional bound on the agents' perception are explored. The results in the different setups show that centralized evolution may lead to somewhat better performance, but at the cost of a large increase in required computations. Also, the decentralized population endogenously learns a more efficient scheme for adaptation. Finally, placing bounds on the agents' perceptive capabilities, which essentially removes reputation as a source of information is shown to have a substantial negative effect on the population.

Scheduled for Session 4.6 Agent-Based Computational Economics - II