Original title: RAPO-RL-TAC: Risk-Aware Partial-Order Reinforcement Learning with Timed Automata Completion for the Interval Job Shop Problem
Pujie Han, Yiheng Liu, Min Huang · September 2026
Scheduling jobs on shared machines is hard when task durations are uncertain ranges rather than exact numbers. Picking an inflexible order too early can cause major delays later. To solve this, researchers designed RAPO-RL-TAC. The method uses reinforcement learning, an artificial intelligence technique that learns by trial and error, to lock in clear scheduling decisions while deliberately leaving risky choices open. Then, a system called timed automata settles those open choices in real time as jobs actually run. Testing on 14 benchmark problems showed that this approach reduced mean midpoint total completion times by 2.92% compared to neighbourhood search and 1.63% compared to a genetic algorithm. It also beat reproduced bee colony methods on 7 of 14 problems, proving that delaying sensitive decisions protects schedules.
You are cooking a dinner with three dishes, and each recipe only gives a range, like 20 to 30 minutes in the oven. If you fix the whole order of cooking before starting, one dish that runs long throws every later dish off. Instead you lock in only the steps you are confident about, keep the uncertain steps flexible, and decide their order when you see how fast the oven is going. In this paper, the job-shop scheduler does the same: it fixes only the safe machine sequences and chooses the uncertain ones as the schedule actually runs.
Manufacturing schedules usually assume each job has one fixed duration, but real machines vary. This work shows that a scheduler which commits only to safe decisions, and defers the rest, can shorten average completion times while keeping schedules stable. For production planners, that means less wasted buffer time and more trustworthy delivery windows.
The abstract doesn't say how the method scales beyond the 14 ORB and LA benchmark instances, whether it transfers to real production settings, or what computational cost it adds. It also only reports comparisons against PNS, GA, and fEABC variants, so relative performance against other methods is not established.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
RAPO-RL-TAC: Risk-Aware Partial-Order Reinforcement Learning with Timed Automata Completion for the Interval Job Shop Problem — Pujie Han, Yiheng Liu, Min Huang — Processes — doi:10.3390/pr14182995
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.