An Action-masking and Self-labeling-based Learning Method for Block Sequencing and Bay Assignment Optimization in Shipbuilding Panel Lines

Abstract

The block sequencing and bay assignment problem in a shipbuilding panel line is a complex combinatorial scheduling problem in which upstream serial processes and downstream parallel bay processes are coupled under multiple operational constraints. In practical panel-line planning, not all field rules have the same level of strictness. Some constraints must be maintained to ensure physical and procedural feasibility, whereas others are relaxable operational constraints that may be relaxed only when no candidate satisfies all rules simultaneously. This study proposes a constraint-priority-based learning method that combines action masking and self-labeling for two-stage panel-line scheduling. The proposed method first constructs a candidate set satisfying hard constraints and then applies relaxable operational constraints according to a predefined priority. If the candidate set becomes empty, only the relaxable operational constraints are gradually relaxed. A pointer-network-based policy selects the next block from the masked candidates, and a rule-based bay-assignment module assigns the selected block to a feasible downstream bay. Multiple rollouts are generated for each instance, and the best rollout is selected as a pseudo-label according to the lexicographic order of constraint noncompliance, makespan, and longi workload balance. Experiments on actual and generated panel-line data show that the proposed method improves makespan and reduces relaxable-constraint noncompliance compared with manual and heuristic schedules while maintaining hard constraints.