DERL_Group

This repo presents the models of the paper Dual-Evolving Reinforcement Learning for Automated Reward Design. The models are trained with our dual-training loop: a reward proposer composes rewards from task primitives, and it learns by policy gradient from the validation performance of the policies that its rewards train.

We release the initial proposer (the Meta-Optimizer model, named after the earlier version of the paper) and the best policy model for each task.

Paper · Project page · Code