This study presents Adaptive Dual-OPtimization with Tree learning Genetic Programming (ADOPT-GP), a dual-loop evolutionary framework that simultaneously discovers symbolic rule structures and calibrates parameters. ADOPT-GP couples adaptive genetic programming with a two-stage parameter tuning process: rapid logistic-regression initialization followed by evolutionary calibration. Across runs, fitness improves by 20%–40% on average. Against a bilevel sequential baseline, ADOPT-GP delivers similar or better accuracy while reducing runtime by over 85%, demonstrating scalability. In a university library evacuation case, it yields diverse, interpretable rules that expose tensions between group cohesion and spatial constraints, supporting context-sensitive behaviors. The approach can advance inverse generative social science (IGSS) by linking behavioral theory with computation and offers practical tools for emergency planning.