Applying reinforcement learning to small scale combat in the real-time strategy game StarCraft:Broodwar

This paper presents an evaluation of the suitability of reinforcement learning (RL) algorithms to perform the task of micro-managing combat units in the commercial real-time strategy (RTS) game StarCraft:Broodwar (SC:BW). The applied techniques are variations of the common Q-learning and Sarsa algorithms, both simple one-step versions as well as more sophisticated versions that use eligibility traces to offset the problem of delayed reward. The aim is the design of an agent that is able to learn in an unsupervised manner in a complex environment, eventually taking over tasks that had previously been performed by non-adaptive, deterministic game AI. The preliminary results presented in this paper show the viability of the RL algorithms at learning the selected task. Depending on whether the focus lies on maximizing the reward or on the speed of learning, among the evaluated algorithms one-step Q-learning and Sarsa(λ) prove best at learning to manage combat units.