Texas A&M Robotics Lab
Agents that learn to bluff
- 3RL papers re-implemented from scratch
- 1v1poker as the research testbed
Iām researching reinforcement-learning-based reward hacking in 1v1 poker, to create interpretable strategies from the agents that win.
To support the work I built Agent Mania, a modular open-source PyTorch RL framework with checkpointing and training dashboards. On top of it I re-implemented the TRPO, DQN and DDQN papers from scratch, which keeps tracking efficient and the logic modular.