Learning GFlowNets from partial episodes for improved convergence and stability

TL;DR

This paper introduces SubTB(λ), a new training objective for GFlowNets inspired by reinforcement learning, which learns from partial episodes to improve convergence speed and stability in complex environments.

Contribution

It proposes SubTB(λ), a novel GFlowNet training method that leverages partial subtrajectories, addressing bias-variance tradeoffs and enabling training in more challenging settings.

Findings

Accelerates convergence in various environments

Enables training with longer action sequences and sparser rewards

Provides insights into gradient dynamics and bias-variance tradeoff

Abstract

Generative flow networks (GFlowNets) are a family of algorithms for training a sequential sampler of discrete objects under an unnormalized target density and have been successfully used for various probabilistic modeling tasks. Existing training objectives for GFlowNets are either local to states or transitions, or propagate a reward signal over an entire sampling trajectory. We argue that these alternatives represent opposite ends of a gradient bias-variance tradeoff and propose a way to exploit this tradeoff to mitigate its harmful effects. Inspired by the TD($\lambda$) algorithm in reinforcement learning, we introduce subtrajectory balance or SubTB($\lambda$), a GFlowNet training objective that can learn from partial action subsequences of varying lengths. We show that SubTB($\lambda$) accelerates sampler convergence in previously studied and new environments and enables training GFlowNets in environments with longer action sequences and sparser reward landscapes than what was possible before. We also perform a comparative analysis of stochastic gradient dynamics, shedding light on the bias-variance tradeoff in GFlowNet training and the advantages of subtrajectory balance.