Junjin Chen Undergraduate Dissertation 2017/18

PLAYING STARCRAFT 2 MINI GAMES WITH DEEP REINFORCEMENT LEARNING

Supervised by E.Vasilaki

Abstract

Inspired by Deep reinforcement learning of Atari games done by Deep Mind which in many games outperforms than human player, this project will implement the deep reinforcement learning on a Real Time Strategy (RTS) game Starcraft2. Applying learning method in RTS games are generally considered challenging since there are usually large action space and large state space in the game as well as large amount of possible strategies for future states, especially in this case where Starcraft2 is a typical RTS game. However, open source from the cooperation between deep mind company and Blizzard entertainment has provided useful and convenient platform for researchers. It includes convolutional layering from the raw data, simplified maps and some example base-agents. Learning methods and the policy-reward functions are only needed to be concern from the researchers. Other researches about other games online also provide the theory of how to apply the learning algorithm into the game. In this article, important terminologies and relevant literature reviews are presented.