Recursive Hyper-Heuristics for the Job Shop Scheduling Problem

Hyper-heuristics are a broad topic that has drawn increasing attention because of its flexibility. This, however, implies that there are diverse models, including selection hyper-heuristics, where the idea is to derive a model that learns when to use each available solver. Nonetheless, such a learning procedure usually proves difficult and leads to non-ideal selections. Hence, in this work, we propose a recursive hyper-heuristic model allowing more complexity within the selection models. Our idea is straightforward: to have a selection hyper-heuristic to select low-level heuristics and lower-level hyper-heuristics. In doing so, one can merge the combined decisions of existing solvers. We test the feasibility of such a model through experiments on the Job Shop Scheduling Problem that cover small and large datasets of previously tailored instances. We found that increasing the order of the model leads to more stable and better-performing approaches. For example, migrating from a second-order hyper-heuristic to a fourth-order hyper-heuristic reduced the makespan by over 6%. Thus, the proposed model seems feasible and should be further tested under more varied scenarios and conditions.