Goal-Space Planning Improves RL for Demand Response Scheduling

A team of researchers has combined Goal-Space Planning (GSP) with Deep Deterministic Policy Gradient (DDPG) to tackle terminal constraints in data-driven demand response scheduling for electrified chemical processes. This innovative method employs learned temporally abstract models across discrete subgoals to enhance value propagation over extended horizons, addressing the credit-assignment difficulties encountered in traditional reinforcement learning. In a simulation of an air separation benchmark, the technique demonstrated improved sample efficiency, met terminal storage requirements, and minimized myopic control actions. This research is available on arXiv (2605.14741) in the Electrical Engineering and Systems Science > Systems and Control category.