WIRL

A reinforcement learning method for long-horizon tasks that treats the multiple rounds of decisions between two milestone outcomes as an execution window and assigns a shared outcome-based reward to the window, striking a compromise between sparse rewards and stepwise dense rewards.