Abstract: Generalisation is important for applying Deep Reinforcement Learning (DRL) algorithms into real world, since real world environments are always changing. A policy suitable for real world scenario must be able to handle these changes. "Generalisation in RL" is a class of problem, and "robust RL" reviewed in my previous talk "Adversarial Attacks and Robustness in DRL" is one of a kind.

Recently, Kirk, Robert, et al. systematically reviewed "Generalisation in RL". They divided generalisation methods into three categories: increasing similarity, handling difference, and RL-specific improvements.

This talk will take their review as the main line, and introduce most generalisation RL works that they cite.