Inferring Behavior-Specific Context Improves Zero-Shot Generalization in Reinforcement Learning

Tidiane CAMARET NDIR, André Biedenkapp, Noor Awad

Keywords: Zero-shot generalization, Contextual reinforcement learning, Meta-learning

TL;DR: This work enables reinforcement learning agents to adapt to new environments without additional training by integrating context learning into policy optimization.

Abstract:

In zero-shot generalization (ZSG) in Reinforcement Learning (RL) agents must adapt to entirely novel environments without additional training. Understanding and utilizing contextual cues, such as the gravity level of the environment, is critical for robust generalization.
We propose to integrate the learning of context representations directly with policy learning. Our algorithm demonstrates improved generalization on various simulated domains, outperforming prior context-learning techniques in zero-shot settings. By jointly learning policy and context, our method develops behavior-specific context representations, enabling adaptation to unseen environments. This approach marks significant progress toward reinforcement learning systems that can generalize across diverse real-world tasks.

Submission Number: 39