A novel machine-learning rolling horizon heuristic for dynamic lot-sizing and job shop scheduling problems

This study introduces an innovative approach to production planning and scheduling under uncertain customer demands by integrating uncertainty directly into the algorithmic framework. We present a novel rolling horizon simulation-based constructive heuristic to minimise total costs, encompassing production, setup, inventory, and backlog costs. We leverage priority rules to enable the method to adapt in real-time to changes. Our contributions include not only a novel heuristic but also the integration of a coevolutionary genetic programming-based hyper-heuristic, significantly improving solution quality and computational efficiency. Compared to other rule-based heuristics, our method consistently outperforms with an average reduction in total costs of 1.78%. Furthermore, it outperforms deterministic and two-stage stochastic programming models within a one-hour time limit, with reductions of 18.20% and 6.53%, respectively. Incorporating mathematical programming models into a rolling horizon scheme led to slightly lower total costs, with an average reduction of 0.13% and 1.16%. However, in less than half an hour, the proposed method reached better results in two out of three cases. The results of an efficiency and robustness analysis highlight the proposed method as a robust solution for dynamic and complex real-world applications.