Towards Deep Optimization in Operations Scheduling

Deep Optimization (DO) is a recently proposed metaheuristic that uses deep neural networks for optimization. It shows promise solving various Combinatorial Optimization (CO) problems; however, it only operates on binary data, requiring problems to be transformed using Quadratic Unconstrained Binary Optimization (QUBO). The study explores the use of DO for the Single-Stage Job Scheduling
Problems (SSJSP), a constrained CO problem. As SSJSP does not naturally fit QUBO, these problems require a workaround to use the number partitioning problem as part of the transformation process and introduce binary and Lagrangian penalty methods to integrate the constraints as part of DO. Performance testing across various parameter settings using generated datasets with controlled slack
demonstrates that the method works well and DO with penalty integration performs satisfactorily on all datasets. Both penalty integration methods are found to be effective in guiding DO toward feasible and optimized solutions, however their relative effectiveness varies depending on the specific scenario.