Seminar: Deep Reinforcement Learning for Real-time Strategy Game Artificial Intelligence
M.Sc. Thesis Proposal
Supervisor: Dr. David Churchill
Deep Reinforcement Learning for Real-time Strategy Game Artificial Intelligence
Department of Computer Science
Friday, November 23, 2018, 2:15p.m., Room EN 2022
Games have long been considered important Artificial Intelligence (AI) testbeds, from Chess to Go and now to more complex video games such as the real-time strategy (RTS) game StarCraft II. RTS games are military simulations played on 2D maps with up to hundreds of units to control and include hidden information or “fog of war”. The real-time nature of RTS games and incomplete information is closer to real-world situations than classic board games, making them good environments for AI advancement. The current state-of-the-art in RTS AI involves hand-crafted strategies using expert human knowledge mixed with various AI techniques such as heuristic search. These techniques are not easily transferred to other domains, and skilled human players can defeat these AI players by adapting to their predictable strategies. Reinforcement learning (RL), an AI method in which a system learns to maximize a reward by interacting with its environment, has in recent years been combined with deep neural networks and applied to games with large state spaces, such as Go. The StarCraft II Learning Environment has recently been released, which provides access to feature layers that give an abstraction of what a human player sees and facilitates training deep neural networks. Our goal is to use existing RL methods and supervised learning from human replays to train an agent to play a subset of the full game of StarCraft II.