Process optimization via neural network metamodeling

Abstract

Listen

Translate article

Queuing systems are often used to model manufacturing or communication systems. Simulation models have been the most accurate models of such systems, but when optimizing the systems it is often impractical to simulate all possible configurations to identify the “best” choice. In the absence of mathematical approximations, one may turn to simulation metamodeling. Traditional metamodeling schemes would involve a simulation model of the entire system run over some subset of the possible configurations. A regression model is then typically fit to the results of these runs and used to predict results for the many configurations that were not simulated. Our approach is to develop an artificial neural network (ANN) metamodel of components of the entire system rather than of the entire system. The entire system is then modeled by interconnecting the ANN metamodels. Specifically, we propose the following approach. An ANN is trained via simulation to act as a single “generic” queuing node. The ANN accepts input data on inter-arrival and service distributions in the form of histograms and generates a departure distribution and a performance measure. A network of queues is then modeled by interconnecting copies of the generic queuing node. To illustrate the process, a simple manufacturing system is modeled. The ANN network model is used to select the buffer sizes in front of each process that provide a desired mix of product throughput rates, while minimizing average system sojourn time.