Tabular Reinforcement Learning in Real-Time Strategy Games via Options
Anderson Rocha Tavares, Luiz Chaimowicz · 2018
Real-Time Strategy (RTS) games are complex domains with huge state and action spaces. In such games, humans usually pursue long-term plans, which take long sequences of actions to achieve. In this work, we implement this behavior by reasoning over options, which are temporally-extended actions in Markov Decision Processes. Our options are defined with the aid of a state aggregation scheme and a portfolio of game-playing algorithms. Experimentally, we show that our approach leverages the capabilities of traditional reinforcement-learning techniques, which become competitive against state-of-the-art search methods in μRTS.