Abstract: Task and personnel scheduling problems are critical to production management. In the condition of time-of-use electricity tariffs, it is difficult to strike a balance between manufacturing and labor costs: electricity prices are lower at night but the salary for working-at-night is higher, while hourly wages are lower during the day but electricity prices are higher. If the two problems are jointly modeled, the resulting large-scale model is not easy to solve. In applications, it is usually addressed in the way that the task scheduling is firstly solved, and then the personnel scheduling is determined afterwards; however, it is difficult to guarantee a low-cost solution. To resolve this issue, this paper proposes a surrogate model method, which uses: 1) multiple sets of relatively optimal feasible solutions of the two sub-problems generated by genetic algorithm, as training samples; and 2) BP neural network, deep learning and broad learning systems to respectively fit the surrogate model for the combination problem. Then, BFGS algorithm is used to search for optima. The proposed adaptive sampling algorithm can effectively simplify the problem in terms of dimensions as the number of jobs and processes increases. Results show that the new method can obtain significantly better results than those obtained by genetic algorithm. In addition, it can save up to 11.9 % of the total cost of electricity and manpower for enterprises.

Key words: surrogate model, simulation-based optimization, broad learning system, variable metric method, adaptive sampling