Undergraduate student Jonas Buro was awarded the NSERC Undergraduate Student Research Award (USRA) to work on Deep Reinforcement Learning in Abstract Strategy Games with the GAIDG Lab.

Recent progress in reinforcement learning and blended approaches to game playing which use reinforcement learning has led to superhuman performance in game-playing artificial intelligence (DOTA2, Starcraft, Chess, GO, Shogi). The most state-of-the-art of these methods use model-free reinforcement learning that has no need to know apriori the rules of the game to be played. Reinforcement learning seeks to model problems as an agent observing and receiving awards in some environment assuming the underlying model is a Markov Decision Process (MDP), or Markov Game in the multi-agent case. In this way, model-free means we may not know something about the MDP, e.g. the transition table, or the reward signal, and thus need to learn through trial and error. Learning arbitrary abstract strategy games, like chess, in this way is particularly difficult. The agent must learn the rules, the moves, the tactics, and strategies that lead to success. This type of problem-solving has real-world applications in robotics, decision making, resource allocation, industrial optimization, and more. That is, learning to maximize future rewards through trial and error when significant useful information is missing or unavailable has applications in many domains outside of games.