Deep learning i Starcraft 2 : Autoencoders för att förbättra end-to-end learning

• Main Authors: Frick, Victor, Mattsson, Kristoffer
• Format: Others
• Language: Swedish
• Published: 2018
• Subjects: Computer and Information Sciences; Data- och informationsvetenskap
• Online Access: http://urn.kb.se/resolve?urn=urn:nbn:se:hb:diva-25531

A problem for deep learning has for a long time been complex environments. Recently end-to-end learning agents have been used to master Atari games by processing raw pixel data into features and using these features to make good decisions. This method has not had the same success in the real time strategy game Starcraft 2 and the authors of this paper decided to investigate the possibility of using autoencoders to train feature extractors and thereby improving the rate of learning for reinforcement learning agents. Asynchronous Advantage Actor Critic agents are used for investigating the difference and the use of the PySC2 API enables tests in the Starcraft 2 environment. The results show that the agents need more training to be able to evaluate the pros and cons of an pretrained feature extractor. However, the training time of the autoencoder was short and if it turns out to improve the performance the authors see no arguments not to use an autoencoder to pretrain a feature extractor in Starcraft 2.