Optimization of Neural Networks with Response Surface Methodology: Prediction of Cigarette Pressure Drop

Abstract

Neural network is an artificial intelligence technique providing successful results in solving many prediction problems. There are many factors affecting the predictive performance of a neural network such as learning rate, momentum rate, number of neurons etc. In practice, trial-error method is used in the selection of these parameters. However, this approach is a time-consuming process and can only measure the effect of change in one parameter on performance at a time. In recent years, alternative methods such as experimental design, genetic algorithm, simulated annealing have been used to find the optimum neural network topology. In this study, response surface method, which is one of the most common design of experiment, was used to determine the neural network topology that would provide the highest predictive performance of the pressure drop that is a quality parameter in tobacco industry. The parameters affecting the pressure drop parameter are considered as circumference, total weight and ventilation. In this context, number of hidden neurons, learning rate, momentum rate and stop criteria were identified as the experimental factors and the combination that gives the lowest mean absolute deviation has been proposed as neural network model to predict pressure drop parameter. As a real life application, thousand cigarette samples have been processed by multilayer perceptron. Findings revealed that epoch size, learning rate, number of hidden neurons and stop criteria have significant linear impact on mean absolute deviation of neural network. Optimum neural network design has been obtained to predict pressure drop parameter.