Design of Experiments for Artificial Neural Network Architecture

Artificial Neural Networks are nonlinear multivariable models that link input-output data as an adaptive learning process meant to mimic pattern recognition in human brains. The construction of neural network models has been described as an artform owing to the number of input nodes, the number of hidden layers, the number of epochs, the learning rate, and the choice of activation functions. The use of design of experiments to identify appropriate factor settings that increase neural network accuracy is an attractive solution as opposed to an ad-hoc or naïve approach. This work develops a simulation metamodel that accurately represents the input-output relationship of a discrete-event model using an artificial neural network that has been constructed using design of experiments to reduce predictive loss and increase accuracy of the network. Using the output of the discrete-event simulation as labeled ground truth data allows backpropagation of the neural network where different architectures are evaluated using design of experiments to select hyperparameter settings. The use of simulation metamodels as an analytic solution to discrete-event models reduces the computational cost and runtime associated with execution.