Flexible job-shop scheduling is a hard combinatorial problem, and neural solvers have become a popular approach. These solvers typically make decisions one operation-machine assignment at a time, refreshing a graph representation of the problem and updating feasibility after every step.
The authors of a new arXiv preprint argue that this is overkill. Only the feasibility check must be exact at each step; the graph representation used to guide the next decision does not need to be recomputed from scratch every time. Their method, CoRS, reuses composite interval states to maintain the information needed for fast, accurate decisions while updating feasibility exactly.
This separation of concerns could make neural FJSP solvers faster in practice, since the expensive graph refresh is a bottleneck. The abstract does not report numerical results in the excerpt, so the speedup remains to be quantified, but the idea targets a clear inefficiency in current approaches.