A Deeper Insight of Neural Network Spatial Interaction Model's Performance Trained with Different Training Algorithms

Abstract

Neural Network (NN) expansion can be categorized into three crucial stages, namely before 1960, 1960-1986, and after 1986 periods. The era before 1960 is when NN was discovered and developed; the next period saw a decline in NN research, while the last period has witnessed advanced developments in NN, especially in improvement of the training algorithms. Therefore, training algorithm is considered as the key in the successfulness of NN application. However, inappropriate choice of training algorithms may lead to poor NN performance. This paper reports the performance of NN models for spatial interaction modelling trained by three different algorithms and discusses some fundamental issues such as the training error and gradient, connection weight update, mapping output, performance consistency, and over fitting. These issues are rarely discussed in previous studies. Findings from this study are expected can assist the transport modeller in using NN as a robust and sound modelling tool.