Flow Matching for Offline Reinforcement Learning with Discrete Actions
The Q-weighted discrete flow matching (QDFM) framework extends flow matching to handle discrete action spaces with multiple objectives. It replaces continuous vector fields with a Continuous-Time Markov Chain (CTMC) approach, using a Q-weighted conditional DFM loss. This method recovers optimal policies from multimodal offline data and allows for generating actions based on changing preferences without retraining.