NeurIPS 2020

Predictive Information Accelerates Learning in RL

Meta Review

This paper experimentally checks the hypothesis that capturing the predictive information is useful in RL. The novelty of the proposed auxiliary task lies in the fact that it learns a *compressed* representation of the predictive information. The experiments are convincing in showing the improvement of PI-SAC over SAC (with or without data augmentation) and over other approaches using auxiliary tasks.