AGGREGATE PRODUCTION PLANNING INTEGRATING CLEARING FUNCTION USING NONLINEAR FUNCTION OPTIMIZATION AND SUPERVISED MACHINE LEARNING

Abstract

Infeasibility in aggregate production planning in material requirements planning (MRP) is mainly caused by the assumption that lead time is constant and load-independent. However, defining a load-dependent lead time may create a planning circularity problem since lead time depends on resource utilization, which is determined by the aggregate planning models' assignment of jobs to resources. One approach to solving this issue is incorporating the clearing function (CF) into an aggregate planning production model.

Recently, a Fixed-Point Release (FPR) model for the closed queuing network was proposed, and promising results were obtained when approximating the CF for the system instead of at the bottleneck station. However, the model's accuracy is a trade-off between the step size selection and computation time. The first objective of this thesis is to develop a nonlinear model by directly integrating the analytic clearing function for an open queuing network into the aggregate production planning called the Clearing Function for an Open Queuing Network (QCF-O). The model's utility is illustrated with a numerical case study, verified by a discrete simulation model, and compared with the current FPR models.

The second objective of this thesis is to develop a framework that integrates supervised machine learning (SML) for aggregate production planning to avoid circularity for closed networks. In this framework, the trained artificial neural network (ANN) and random forest (RF) models are used as inverse maps to estimate discrete values of the CF for a Closed Queuing Network. This discretized CF is then used in place of the FPR model of Rayan et al. The framework is also analyzed for its performance in terms of changes in processing time and machine failures.

Finally, the thesis nonlinear model solution approach for the first objective in the thesis is replicated by using NOMAD (Nonlinear Optimization by Mesh Adaptive Direct Search), a nonlinear blackbox optimization using Mesh Adaptive Direct Search to present an approach which can eventually be based on data-driven model-free estimations of the CF using ML.