search

EmptyJackson / groove

Official implementation of the NeurIPS 2023 paper "Discovering General Reinforcement Learning Algorithms with Adversarial Environment Design"
Apache License 2.0
22 stars 4 forks source link

Code for Generator #1

Closed youwyu closed 3 months ago

youwyu commented 3 months ago

May I ask where is the code for Generator in the paper? "the level generator is a generative model that is optimized to produce regret-maximizing levels"

EmptyJackson commented 3 months ago

The generative model mentioned here is from the definition of Dual Curriculum Design, which describes a class of UED methods consisting of a generator and a curator. We base GROOVE on Prioritized Level Replay, which doesn't optimise a generator and instead curates randomly-sampled levels (see end of Section 3.3).

Hope this helps!