Jesse van Remmerden

Jesse van Remmerden

PhD candidate at TU/e. Reinforcement learning, diffusion models and LLMs for scheduling, routing and maintenance planning.

Generalizing Beyond Suboptimality: Offline RL Learns Effective Scheduling Through Random Solutions

Date:

I presented CDQAC in the Smart Industry track at NWO ICT.OPEN 2026, held at De Doelen in Rotterdam on 15–16 April 2026 under the theme The Art of ICT. The track was chaired by Jurjen Helmus (HvA) and John Bolte (HHS).

The talk made the case that the simulator requirement, long treated as the price of admission for deep RL in scheduling, is not actually necessary — and that the data such a method learns from need not be any good. CDQAC learns scheduling policies entirely offline from previously executed schedules, reaches state-of-the-art performance from 10–25 training instances, and performs best when trained on solutions generated by a random heuristic. The underlying reason is that offline performance tracks how much of the decision space the dataset covers rather than how good its solutions are, a relationship that holds at a rank correlation of −0.95.

For an industrial audience this reverses the usual order of operations: rather than first building a faithful simulator or curating a set of expert schedules, it is enough to log the decisions a plant already makes — or even to sample them at random.

Links