Daegyeong Roh

PAMD: Structured Adaptive Distances for Bisimulation Representations in Visual Reinforcement Learning

Daegyeong Roh, Juho Bae, Han-Lim Choi

Many visual reinforcement learning (RL) algorithms learn representations by matching latent distances to a behavioral distance induced by reward and transition similarity. In practice, the choice of the latent distance can strongly affect performance: using a fixed, pre-specified…Read more