ivo.rl

Reinforcement learning, mostly offline and on real logged data rather than simulators. Distribution shift, off-policy evaluation variance, and the gap between a policy that looks good in evaluation and one that survives deployment.