Instance based Generalization in Reinforcement Learning

TL;DR

This paper investigates how instance reuse affects reinforcement learning generalization, providing theoretical bounds and proposing a shared belief representation method to improve performance on unseen environments.

Contribution

It formalizes the impact of instance reuse on RL dynamics, introduces a belief ensemble approach, and validates these insights through experiments on CoinRun.

Findings

Instance reuse significantly alters effective Markov dynamics during training.

Maximizing rewards can lead to instance-specific policies that hinder generalization.

Shared belief representations improve performance on unseen levels.

Abstract

Agents trained via deep reinforcement learning (RL) routinely fail to generalize to unseen environments, even when these share the same underlying dynamics as the training levels. Understanding the generalization properties of RL is one of the challenges of modern machine learning. Towards this goal, we analyze policy learning in the context of Partially Observable Markov Decision Processes (POMDPs) and formalize the dynamics of training levels as instances. We prove that, independently of the exploration strategy, reusing instances introduces significant changes on the effective Markov dynamics the agent observes during training. Maximizing expected rewards impacts the learned belief state of the agent by inducing undesired instance specific speedrunning policies instead of generalizeable ones, which are suboptimal on the training set. We provide generalization bounds to the value gap in train and test environments based on the number of training instances, and use insights based on these to improve performance on unseen levels. We propose training a shared belief representation over an ensemble of specialized policies, from which we compute a consensus policy that is used for data collection, disallowing instance specific exploitation. We experimentally validate our theory, observations, and the proposed computational solution over the CoinRun benchmark.