We Developed Deep Q-Networks (DQNs) with transfer learning to adapt knowledge from single-player to two-player game environments, resulting in improved training efficiency and performance across ten Atari 2600 games.
See our paper.
See the source code.
projects
- Single-player to Two-player Knowledge Transfer in Atari 2600 games
- Random Network Distillation in Pytorch
I developed a pytorch version of OpenAI’s RND (Random Network Distillation with Proximal Policy Optimization) and trained it on Montezuma’s Revenge to address the challenge of sparse rewards.See the source code.
- Adaptive MCTS with TD Learning in miniXCOM
We dsigned and implemented an adaptive Monte Carlo tree search algorithm enhanced with temporal difference learning, to improve performance without requiring pre-training in the simplified version of the game XCOM.
See our paper.
See the source code.