Automatic Risk Adaptation in Distributional Reinforcement Learning
In a new preprint, we explored the use of the parametric uncertainty of an agent as a way to learn in risk-sensitive environments. We show improved robustness to changing environment dynamics and lower failure rates in hard locomotion environments.
Recently, my co-authors and I extended our method TOAD-GAN to generate Minecraft structures in arbitrary sizes from a given example snippet. Our method World-GAN has been accepted at the Conference on Games 2021.
Our work TOAD-GAN got published at the AIIDE 2020 conference and won the Best Student Paper Award.