Reinforcement Learning Agents with Generalizing Behavior

摘要

We explore the generality of Reinforcement Learning (RL) agents on unseen environment configurations by analyzing
the behavior of an agent tasked with traversing a graph based environment from a starting position to a goal position. We find that training on a single task is likely to result in inflexible policies that do not respond well to change. Instead, training on a wide variety of scenarios offers the best chance of developing a flexible policy, at the expense of increased training difficulty.