Modeling non-stationary opponents

This paper studies repeated interactions between an agent and an unknown opponent that changes its strategy over time. We propose a framework for learning switching nonstationary strategies. The approach uses decision trees to learn the most up to date opponent’s strategy. Then, the agent’s strategy is computed by transforming the tree into a Markov Decision Process (MDP), whose solution dictates the optimal way of playing against the learned strategy. The agent’s learnt model is continuously re-evaluated to assess strategy switches. Our method detects such strategy switches by measuring tree similarities, and reveals whether the opponent has changed its strategy and a new model has to be learned. We evaluated the proposed approach in the iterated prisoner’s dilemma, outperforming common strategies against stationary and non-stationary opponents.