Neuro-Symbolic Reinforcement Learning for Multi-Category Construction Relationship with Reasoning and Sequence Planning

Abstract: Construction planning relies on tacit knowledge that is rarely documented, leading to recurring knowledge loss and limited standardization. This paper proposes a neuro symbolic reinforcement learning framework for multi category construction relationship reasoning that formalizes construction execution relationships as explicit constraints. The framework combines synthetic Work Breakdown Structure generation, domain constrained relationship inference using large language model based reasoning, and policy learning via Group Relative Policy Optimization with symbolic rewards to produce explainable construction relationship graphs at inference time. Domain constraints serve a dual role by grounding reasoning during data generation and providing verifiable reward signals during optimization. Experimental results demonstrate stable learning behavior, with structural validity and reasoning quality exceeding 75 percent across categories, bounded policy divergence with KL approximately 0.22, and a truncation ratio reduced to 13.5 percent, supporting more efficient, consistent, and reusable automated construction planning across projects and organizational contexts.