Computational Neural Networks: An Attractive Class of Mathematical Models for Transportation Research

Most transportation research techniques and methods currently in use were developed in the 1960s and 1970s, i.e. in an era of scarce computing power and small data sets. Their implementations take only limited advantage of the data storage and retrieval capabilities of modem computational techniques, and basically ignore both the emerging new era of parallel supercomputing and the computational intelligence techniques. This chapter provides a small sized real world example for interregional telecommunication traffic modelling and illustrates its superiority compared with the standard statistical benchmark. It considers some fundamental characteristics of these computational neural networks (CNN). The chapter focuses on feedforward neural networks which provide transportation researchers with a novel and extremely useful class of mathematical tools. It deals with supervised training of such networks and reviews some powerful (local) optimization techniques. Multilayer feedforward CNNs such as perceptrons and radial basis function networks have emerged as attractive class of CNNs based upon sound theoretical concepts.