Dual-Evolving Reinforcement Learning for Automated Reward Design
Dual-Evolving Reinforcement Learning (DERL) automatically discovers optimal reward structures by treating meta-optimization as an RL problem. It uses a bi-level framework where an outer loop proposes rewards and an inner policy learns from them. The proposer learns by observing the inner policy's performance, effectively capturing a 'meta-gradient' of task success to find highly effective reward signals.