Decision Transformer for Atari Games

Project Link: https://github.com/Jaron-U/decision-transformer-atari

This study applies Decision Transformers (DT), which reimagine reinforcement learning (RL) as sequence modeling using Transformer architecture, to the Atari game Pong. Traditional RL methods, often hindered by complexities and sparse rewards, are outperformed by DTs that utilize complete trajectories of states, actions, and rewards for enhanced learning. We have advanced the DT framework by integrating a loss-decay algorithm and conducting an ablation study on context length, which significantly improves performance over conventional offline RL algorithms. Furthermore, we explored the effects of model pruning on DT, aiming to optimize efficiency without compromising performance. These adaptations not only enhance strategic gameplay in Pong but also underscore the broader applicability of DTs in addressing complex AI challenges.

This is the ppt of the project: If the embedded PDF below does not load, you can download it here.