Optimization model selection for simulation-based approximate dynamic programming approaches in semiconductor manufacturing operations

Guided by Little´s law, decision and control models for operations in reentrant line manufacturing (RLM) systems are commonly set up to minimize the total work-in-process (WIP), which in turn indirectly minimizes cycle time (CT). By viewing the problem fundamentally differently, we re-formulate it as one that seeks to select the best cost function leading to optimal cycle times. We present the details and results of an extended simulation study, based on a benchmark problem, using a simulation-based approximate dynamic programming method, with a newly proposed extended actor-critic architecture. Our results support the idea that a Markov decision process modeling approach can be used as a flexible platform to explore different cost formulations, leading to a selection of an optimal cost and model to optimize cycle time directly.