Reliability-Guided Allocation of Genetic Algorithm Tuning Effort in Permutation Flow Shop Scheduling: A Prospective Validation Study

Genetic algorithms (GAs) are widely used for permutation flow shop scheduling, but reusing calibrated parameters at a new problem size raises a practical question: how much additional tuning is justified? This study evaluates a reliability-guided rule for allocating that effort. Nine Taillard size groups provide a transferable parameter center and an empirical transfer-reliability index based on held-out same-size instances. For a new size, interpolated reliability determines search breadth and replication effort: GREEN screens the center and available ±1 axial neighbors, AMBER a local Cartesian neighborhood, and RED the full grid, followed by confirmation of leading candidates. The prospective protocol was fixed before outcomes from 24 off-grid targets were observed. At the prespecified 0.25-percentage-point tolerance, both the policy and direct center reuse were within tolerance on 21/24 targets; statistical superiority over direct reuse was not established, although intervention-only analyses favored guided tuning under stricter and continuous-loss summaries. A staged implementation required 3767 tuning runs, versus 30,000 for universal 125 × 10 retuning and 3768 for a reliability-blind budget-matched tuner. Under an interpretive ±0.05% band, fresh validation yielded 9 guided-better, 14 practically tied, and 1 blind-better target. These results support historical transfer reliability as a risk-ranking signal for selective tuning without implying universal superiority.