Routing under machine breakdowns: a benchmark of dispatching rules, bandits, and reinforcement learning for multi-server flow shops

Reinforcement learning is increasingly recommended for routing decisions in flow-shop manufacturing, yet head-to-head comparisons against the dispatching rules it would replace are often weak or absent. This study addresses that gap for machine breakdowns. We benchmark five dispatching rules, two online bandit methods, and Proximal Policy Optimisation (PPO) on a multi-server flow shop simulator with opt-in non-stationarity mechanisms. Five breakdown severities plus a stationary baseline are tested: 4,200 evaluation episodes for rules and bandits across two industrial testbeds and 2,750 for PPO on electronics. On the primary electronics testbed, ShortestQueue and the Hybrid Bandit-Queue (HBQ), a diagnostic queue-aware bandit, lead decisively at mild and moderate severity. Under the heaviest disruption tested they converge to within 3% of the other rules and bandits. On these testbeds availability information confers no measurable benefit over the queue signal; on ten public Carlier-Néron topologies HBQ is the most robust method under heavy breakdowns, consistent with its availability gate. PPO trained under common public defaults performs far worse. Re-training under a corrected protocol (higher discount factor, held-out evaluation-cost checkpointing, throughput-aware reward) narrows the cost-per-unit gap to ShortestQueue on the breakdown grid from 77–137% to 38–72%, without closing it; the stationary gap is larger. All five trained policies remain costlier than ShortestQueue in all 30 cells (one-sided configuration-level signed-rank p = 0.016), and per-seed analysis identifies a bimodal training failure mode that the corrected protocol mitigates without eliminating. We conclude with a method-selection guide bounded to the disruption regimes tested.