| The Prisoner's Dilemma (PD) has been discussed extensively to model the conflict between competition and cooperation, or between individual and collective rationality [1][8][9][10][11]. The machine learning community has an interest in the iterated PD (IPD) game, but has special interest in the behavior of the individual in the IPD game.;A family of estimate-based strategies, which are based on the pursuit scheme and interconnected Learning Automaton (LA) structure, is developed in this Thesis. These achieve a high performance in playing the IPD game in Nonstationary Environments. Simulation results show that, our proposed scheme, the IPPP (Interconnected Learning Automata with the Preceding Penalty Limit Pursuit), achieves 4.86% lower costs than the SLASH (Stochastic Learning Automata with States of History) strategy, which is the most efficient learning strategy reported in the literature. The advantages of IPPP are its quick detection of strategy switches, fast convergence, and its good generalization capability. |
