Opponent awareness at all levels of the multiagent reinforcement learning stack

Multiagent Reinforcement Learning (MARL) has experienced numerous high profile successes in recent years in terms of generating superhuman gameplaying agents for a wide variety of videogames. Despite these successes, MARL techniques have failed to be adopted by game developers as a useful tool to be used when developing their games, often citing the high computational cost associated with training agents alongside the difficulty of understanding and evaluating MARL methods as the two main obstacles. This thesis attempts to close this gap by introducing an informative modular abstraction under which any Reinforcement Learning (RL) training pipeline can be studied. This is defined as the MARL stack, which explicitly expresses any MARL pipeline as an environment where agents equipped with learning algorithms train via simulated experience as orchestrated by a training scheme. Within the context of 2-player zero-sum games, different approaches at granting opponent awareness at all levels of the proposed MARL stack are explored in broad study of the field. At the level of training schemes, a grouping generalization over many modern MARL training schemes is introduced under a unified framework. Empirical results are shown which demonstrate that the decision over which sequence of opponents a learning agent will face during training greatly affects learning dynamics. At the agent level, the introduction of opponent modelling in state-of-the art algorithms is explored as a way of generating targeted best responses towards opponents encountered during training, improving upon the sample efficiency of these methods. At the environment level the use of MARL as a game design tool is explored by using MARL trained agents as metagame evaluators inside an automated process of game balancing.