Evolving Curricula with Regret-Based Environment Design

#1

Post by minch » Fri, Mar 04, 2022, 1:14 PM UTC

Evolving Curricula with Regret-Based Environment Design
accelagent.github.io

Re: Evolving Curricula with Regret-Based Environment Design

#2

Post by aiaffinity » Sat, Mar 05, 2022, 12:17 AM UTC

Very interesting work and an amazing demo!! Btw, a very basic question: in the demo is there any curricula/training process being shown, or is it basically the final policies learned by each of the four competing approaches?

(ps: I co-wrote a short elementary paper on auto curricula design for RL in 2017 [0])

[0]: https://arxiv.org/abs/1703.07853

Re: Evolving Curricula with Regret-Based Environment Design

#3

Post by minch » Sat, Mar 05, 2022, 12:04 PM UTC

aiaffinity wrote: post #2

Very interesting work and an amazing demo!! Btw, a very basic question: in the demo is there any curricula/training process being shown, or is it basically the final policies learned by each of the four competing approaches? (ps: I co-wrote a short elementary paper on auto curricula design for RL in 2017 [0]) [0]: https://arxiv.org/abs/1703.07853

Thanks! The demo just shows the final agents after training (30K gradient updates). Interesting work re the reward maximizing curricula. I have not seen this before, so thanks for the pointer.