Steel plate allocation via heuristic-guided reinforcement learning

Abstract

Order allocation in steel manufacturing constitutes a large-scale NP-hard combinatorial optimization problem. The task assigns heterogeneous subplates to motherplates to minimize material waste, directly influencing production cost. Existing approaches rely on simplified assumptions or predefined cutting patterns, which limit scalability and adaptability in dynamic production scenarios. We reformulate subplates allocation as a sequential decision problem and propose a framework integrating heuristic primitives with reinforcement learning. The framework defines a structured action space using a finite set of parameterized local transformations, avoiding exhaustive enumeration of combinatorial configurations. A two-stage learning paradigm is adopted, where heuristic-guided evaluation provides stable decision priors, and a reinforcement learning policy progressively refines decisions through long-horizon interaction. Experimental results on large-scale industrial datasets demonstrate that the proposed framework attains the theoretical minimum number of motherplates in most instances. Moreover, it concentrates 95 % of residual material onto a single target plate in over 55 % of cases, with the utilization of contributing motherplates exceeding 99.8 % on benchmarks. The framework generalizes well across diverse industrial order structures and satisfies real-time production requirements with second-level response times.