This repository contains the official implementation of the paper:
Enhancing Two-Player Performance Through Single-Player Knowledge Transfer:An Empirical Study on Atari 2600 Games
Authors: Kimiya Saadat, Richard Zhao
Published in Proceedings of the Twentieth AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment (AIIDE2024).
This is an implementation of two player training in ten Atari 2600 environments by using the Atari RAM as the input, using self-play. There is an option to use an already trained single-player network and only train the final layer to leverage the knowledge gained in single-player Atari environments. Our results show that by using transfer from single-player instead of training from scratch, the overall average score gained by the agent improves in some of the games.