Path-Coupled Bellman Flows for Distributional Reinforcement Learning
Path-Coupled Bellman Flows (PCBF) is a continuous-time DRL method that uses flow matching to learn return distributions by enforcing consistency between the current and successor paths of the Bellman equation. It introduces a $\lambda$-parameterized target that allows users to trade off bias for variance reduction, leading to improved distributional fidelity and training stability.