Reinforcement Learning for Build-Order Production in StarCraft II

Abstract

StarCraft II is one of the most popular real-time strategy games and has become an important benchmark for AI research as it provides a complex environment with numerous challenges. The build order problem is one of the key challenges which concern the order and type of buildings and units to produce based on current game situation. In contrast to existing hand-craft methods, we propose two reinforcement learning based models: Neural Network Fitted Q-Learning (NNFQ) and Convolutional Neural Network Fitted Q-Learning (CNNFQ). NNFQ and CNNFQ have been applied into a simple bot for fighting against the enemy race. Experimental results show that both these two models are capable of finding the most effective production sequence to defeat the opponent.