A Survey of Zero-shot Generalisation in Deep Reinforcement Learning

The study of zero-shot generalisation (ZSG) in deep Reinforcement Learning (RL) aims to produce RL algorithms whose policies generalise well novel unseen situations at deployment time, avoiding overfitting their training environments. Tackling this is vital if we are deploy reinforcement learning real world scenarios, where the environment will be diverse, dynamic and unpredictable. This survey an overview nascent field. We rely on a unifying formalism terminology for discussing different ZSG problems, building upon previous works. go categorise existing benchmarks ZSG, as current methods tackling these problems. Finally, provide critical discussion state field, including recommendations future work. Among other conclusions, argue that taking purely procedural content generation approach benchmark design not conducive progress suggest fast online adaptation RL-specific problems some areas work recommend underexplored problem settings such offline reward-function variation.