Because of the combination of classification, association, adaptation, and pattern recognition capabilities, neural networks are shown to be suitable for solving problems in production planning with uncertain and non-stationary demand. We demonstrate that a properly designed and trained multi-layered perceptron outperforms traditional algorithms for the rolling horizon version of the dynamic lotsizing problem. Formal arguments are supported by numerical experiments.
Keywords: Lotsizing, Multi-Layered Perceptrons, Neural Networks, Pattern Recognition,
Production Planning, Uncertainty.