Planning or Reinforcement Learning for Multi-Asset Maintenance: Reliability and Cost

Summary

This study compares planning and reinforcement learning (RL) for maintenance scheduling across multiple interacting bearing assets using run-to-failure data. Both approaches are evaluated under different penalties for equipment failure while balancing preventive maintenance and operating cost. Planning treats reliability as a hard constraint and consistently produces policies with zero failures; its total cost changes little as the failure penalty increases. RL agents optimize expected cost instead, often accepting occasional failures in exchange for less preventive maintenance when penalties are low. They continue to produce non-zero failures in some settings even when failure penalties are high, although they can achieve lower costs in low-penalty regimes. The study also tests reward shaping and action masking as lightweight ways to encourage more reliable RL behavior. The authors suggest using planning when strict reliability and short deployment horizons are priorities, and RL when limited failures are acceptable and long-run operating efficiency matters. They present the controlled benchmark, which standardizes the environment, cost model, and evaluation, as a reusable basis for comparisons in other maintenance settings.