Literature
Display
← All papers

Goal Achievement Guided Exploitation: A Principled Performance-Based Scheduling Framework for Reinforcement Learning

Shengchao Yan · Baohe Zhang · Chenguang Huang · Joschka Boedecker · Wolfram Burgard

· Transactions on Machine Learning Research ↗

Reading note

GAGE adjusts exploration according to achieved performance, constraining action-distribution spread or discrete-action logits as goals become easier to reach. The paper compares this scheduling principle with entropy-based approaches. Its robot-control experiments examine when progress itself can provide a useful exploration signal.

Concept demonstration

Open image ↗