Opponent Modelling in Multi-Agent Systems

Reinforcement Learning (RL) formalises a problem where an intelligent agent needs to learn and achieve certain goals by maximising a long-term return in an environment. Multi-agent reinforcement learning (MARL) extends traditional RL to multiple agents. Many RL algorithms lose convergence guarantee in non-stationary environments due to the adaptive opponents. Partial observation caused by agents’ different private observations introduces high variance during the training which exacerbates the data inefficiency. In MARL, training an agent to perform well against a set of opponents often leads to bad performance against another set of opponents. Non-stationarity, partial observation and unclear learning objective are three critical problems in MARL which hinder agents’ learning and they all share a cause which is the lack of knowledge of the other agents. Therefore, in this thesis, we propose to solve these problems with opponent modelling methods. We tailor our solutions by combining opponent modelling with other techniques according to the characteristics of problems we face. Specifically, we first propose ROMMEO, an algorithm inspired by Bayesian inference, as a solution to alleviate the non-stationarity in cooperative games. Then we study the partial observation problem caused by agents’ private observation and design an implicit communication training method named PBL. Lastly, we investigate solutions to the non-stationarity and unclear learning objective problems in zero-sum games. We propose a solution named EPSOM which aims for finding safe exploitation strategies to play against non-stationary opponents. We verify our proposed methods by varied experiments and show they can achieve the desired performance. Limitations and future works are discussed in the last chapter of this thesis