Dynamic Dispatching for Re-entrant Production Lines - a Deep Learning Approach

This study presents a dynamic dispatching method for re-entrant production systems by combing dynamic programming (DP) with deep learning. First, we use DP to derive optimal value functions and optimal dispatching policies in a small number of numerical cases. The optimal value functions are then applied to train a deep neural network (DNN). The DNN builds an efficient estimation engine for optimal value functions. Since optimal dispatching decisions can be considered a compressed feature of the optimal value function, the value function estimated by DNN can be quickly mapped to dynamic dispatching policies. The accuracy of DNN dispatching policies is validated by the k-fold cross-validation (k-cv) test in a wide variety of re-entrant systems. Our preliminary investigation shows the potential of DNN in instantaneously generating accurate dynamic dispatching policies.