Care-Net: deep reinforcement learning for distance-decay maximal covering of elderly care facility location considering multi-attribute site suitability

Population aging heightens demand for evidence-based location of elderly care facilities. Classical maximal covering location problem (MCLP) models assume binary coverage, which does not reflect distance decay in accessibility. This study formulates an elderly facility – maximal covering location problem (EF-MCLP) that explicitly optimizes distance-decay-weighted service coverage and travel-distance efficiency. We develop Care-Net, an end-to-end perception – decision solver based on deep reinforcement learning (DRL) to address this NP-hard problem at scale. The perception module encodes multidimensional contextual features from active-aging indicators and space-syntax measures to represent candidate sites during policy learning; the decision module employs a memory-augmented CacheFormer decoder with self-attention and autoregressive selection to capture interdependencies among candidate facilities. To assess the human-centered quality of selected layouts, we further introduce a comprehensive site quality index (CSQI) as a post-optimization evaluation metric. Using a real-world Zhengzhou dataset, Care-Net ran about 12–13× faster than Gurobi with 0.73%–0.98% optimality gaps and provided a 9.5% better solution quality compared to leading DRL methods. On synthetic datasets (including 10,000-node instances), we observed 96–655× runtime speedups. Across both settings, the solver outperformed classical heuristics and DRL baselines. The formulation and solver achieve sub-2% optimality gaps alongside large runtime reductions, enabling large-scale elderly care facility planning with multidimensional contextual representation and comprehensive post-optimization evaluation.