Regularized Offline GFlowNets

Haozhi Wang, yunfeng shao, Jianye HAO, Yinchuan Li

Keywords: Offline Training, Generative Flow Network

TL;DR: A regularized offline training method for Generative Flow Network

Abstract: We propose the regularized offline generative flow networks (RO-GFlowNets) that does not rely on online sampling. Since offline datasets usually cannot cover the entire state space, traditional GFlowNets cannot accurately predict the action sampling probability for each state. To address this problem, RO-GFlowNet aims to minimize the flow matching loss while regularizing the distribution distance of policy and offline datasets. Experimental results show that RO-GFlowNets perform well on offline datasets.