A Neural Network Approach to Performance Analysis of Tandem Lines the Value of Analytical Knowledge

We develop a neural network (NN) metamodeller for efficiently approximating the throughput of different finite-buffer multi-server tandem lines (with varying service rates, number of stations, buffers, and servers). The resulting NN serves as a quick performance evaluation tool and is subsequently used for optimising the tandem-line layout. Specifically, we discuss the optimal allocation of buffer places and optimising service rates where service rates at machines are associated with costs. Our NN metamodelling approach is new as we integrate (biased) analytical queuing knowledge into the training data. The setup and training of the NN metamodeller are discussed in the paper. In particular, we discuss the integration of analytical results from queuing theory. Our numerical studies corroborate the common belief that adding analytical knowledge (in this case from queueing theory) significantly improves the ensuing NN’s prediction power. The framework developed in this paper demonstrates how analytical system knowledge can be integrated with data science in performance evaluation and optimisation. Our message is that even basic NNs, combined with formulae available from OR theory, offer invaluable improvements for building metamodellers in simulation optimisation.