Learning Task Belief Similarity with Latent Dynamics for Meta-Reinforcement Learning

One-line Summary

SimBelief is a new meta-reinforcement learning framework that improves task identification and exploration in sparse reward environments by measuring task belief similarity using latent dynamics.

Plain-language Overview

Meta-reinforcement learning involves teaching an AI to quickly adapt to new tasks by learning from previous experiences. This process can be difficult when the rewards or feedback from the environment are sparse or hard to detect. The new approach, called SimBelief, focuses on identifying similarities between different tasks to improve how the AI explores and learns. By understanding the common features of tasks, SimBelief enables the AI to adapt more efficiently even when rewards are sparse, outperforming existing methods in tests with simulated environments.

Technical Details