Generative Flow Networks (GFlowNets) are a generative framework designed to sample complex structured objects proportionally to a given reward function. Recent work has shown that, in discrete spaces and under appropriate reward shaping, GFlowNets can...
Generative Flow Networks (GFlowNets) are a generative framework designed to sample complex structured objects proportionally to a given reward function. Recent work has shown that, in discrete spaces and under appropriate reward shaping, GFlowNets can be reformulated as maximum-entropy reinforcement learning (MaxEnt RL). This thesis extends this equivalence to continuous state spaces and relaxes a key terminal-state assumption, establishing the connection in a more general setting.
In terms of training, using the Trajectory Balance loss for GFlowNets corresponds to a policy-gradient method with a scalar baseline. However, modern RL typically employs learned value functions as state-dependent baselines to reduce gradient variance and improve training stability. Motivated by this, we show that Soft Actor--Critic (SAC) can serve as an effective training procedure for GFlowNets, leveraging a well-validated MaxEnt RL algorithm supported by extensive follow-up work.
To this end, we study a termination-controlled SAC algorithm in which the policy explicitly chooses termination, and we develop several mitigation strategies to address its early-termination bias. Experiments on synthetic benchmarks demonstrate that the proposed SAC-based methods better approximate the target reward-proportional distribution and reduce premature termination during training.