Monte Carlo Tree Search Strategies in 2-Player Iterated Prisoner Dilemma Games

This study compares a player using Monte Carlo Tree Search (MCTS) against a variety of well-known Prisoner’s Dilemma strategies in 2-player tournaments. The MCTS player has a simple structure and a reasonable computation budget. Nevertheless, it is highly competitive against all tested strategies. As the MCTS player constructs its game tree, it updates the probability of cooperation in response to an opponent’s cooperation or defection. The trajectories of these updatings over the course of play are found to converge toward optimal counter-strategies against the particular opponent being played. In some cases the speed of progress toward an optimal counter strategy hinders the MCTS player