Learning from demonstrations, then from play

Getting a policy off the ground

Learning a game entirely through trial and error can spend a great deal of time on the basics. This project studied how demonstrations and teacher guidance interact with reinforcement learning.

What we built

With SeongJun Kwon and James Edwardsen, I worked on a pipeline that pretrains a policy from teacher-labeled replay states, then continues learning through PPO self-play.

• Behavior-cloning pretraining to give the policy an initial action model.
• A competence-gated curriculum that ties progression to what the policy can do.
• Controlled teacher-guidance ablations, matched checkpoint evaluation and behavior diagnostics.

Beyond a reward curve

The evaluation work asks how an agent behaves and whether comparisons are made under the same conditions. That distinction matters when a training change improves a score without improving the behavior you wanted.