Introducing Symmetries to Black Box Meta Reinforcement Learning

Meta reinforcement learning (RL) attempts to discover new RL algorithms automatically from environment interaction. In so-called black-box approaches, the policy and algorithm are jointly represented by a single neural network. These methods very flexible, but they tend underperform compared human-engineered in terms of generalisation new, unseen environments. this paper, we explore role symmetries meta-generalisation. We show that recent successful meta approach meta-learns an objective for backpropagation-based exhibits certain (specifically reuse rule, invariance input output permutations) not present typical systems. hypothesise these can play important Building off work supervised learning, develop system same symmetries. through careful experimentation incorporating lead with greater ability generalise action & observation spaces, tasks,