Deja Whitmore-Castillo

Game data provides scalable trajectories and verifiable rewards for pretraining agents.

Simulation scales robot training.

Model-based RL wins on sample efficiency; model-free RL handles complex environments better.

Inferring player intent from behavioral data enables AI systems to learn realistic game strategies.

Pairing reward shaping with complementary techniques unlocks sparse-reward learning.

The benchmark that launched deep RL hides what agents actually learn.

Model-based and model-free agents compete to learn efficiently from minimal Atari experience.