ROI-Focus: A Low-Dimensional Action Representation for Resource Allocation in Regional Interception-Protection Tasks

Regional interception-protection tasks require repeated allocation of limited protection resources as incoming pressure advances along predefined corridors. Under low coverage, direct reinforcement-learning policies search an N-dimensional resource-allocation simplex, whereas fixed ROI heuristics rank cells but cannot adapt concentration. We propose ROI-Focus, a low-dimensional action representation that computes task-side return-on-investment (ROI) scores, learns one focus parameter, and decodes it into a feasible regional allocation vector. We instantiate the task with a wargaming transition over map cells; reward is counterfactual task-loss reduction, and ROI scores are decoder inputs. On the modeled instance, ROI-Focus achieves total improvement 39.25 versus 33.52 for Softmax, with high-ROI capture 0.855 while activating only 0.274 of cells. The results show that embedding value ranking into the action representation improves allocation performance and interpretability. Evidence is abstract simulation, not real engagement physics.