Potential for Machine Learning in Optimized Production Planning with Hybrid Simulation

Advanced production planning and scheduling approaches increasingly rely on simulation-based optimization methods. This entails the problem of a high computational effort due to complex models, resulting in limitations for the practical application of otherwise powerful methods. While machine-learning methods offer a potential for performance improvement, approaches for real-life applications with a high complexity are still lacking. This paper explores the potential for machine learning, especially artificial neural networks, used as surrogate models, to improve the performance of a recently developed planning method for real life production planning applications. The simulation considered in this paper is a complex hybrid discrete-continuous model, enabling the method to pursue energy efficiency simultaneously with economic goals, in a complex multi-criteria goal system. The artificial neural network is trained via offline learning and is meant to provide a computationally cheap evaluation of intermediate planning solutions, compiled by an optimization algorithm during an optimization run. The approach is developed and evaluated in a case-study on the food industry, indicating a basic feasibility of the approach but also pointing out necessary future challenges to be solved towards practical applicability.