We propose a novel Dynamic Separator Configuration (DynSep) method that models separator configuration in different rounds as a reinforcement learning task, making decisions based on an incremental triplet graph updated by iteratively added cuts.