A NEURAL-NETWORK APPROACH TO THE ANALYSIS OF CITY SYSTEMS

This study describes a method for analysing systems of cities and for assessing their sensitivity to change, It is based on the premise that the macroscopic appearance of a city is a result of a larger set of u nderlying processes which can be indicated by useful variables. Herein , a neural approach makes use of Kohonen's self-organizing maps (SOM) to create a phenomenological model of the (West) German city system, S OMs can display hidden patterns in input data as well as neighbourhood relations among the cities that make up the system. The 171 measureme nt vectors and 21 variables comprising the city system dataset can be reduced to just four dimensions that represent all relevant features o f the system. The SOM technique permits classification of German citie s into 24 groups that share common characteristics. By inputting a seq uence of small changes to the data about a given city it is possible t o observe whether and how it evolves towards the characteristics of an other group. Some cities (e.g. Frankfurt, Stuttgart) are relatively in sensitive to these data manipulations, whereas others respond quickly (e,g, Nurnberg), It is believed that the former are core representativ es of discrete city types. With further refinement and broader applica tion to global datasets, this technique may be useful for identifying cities that are susceptible to perturbations of human-nature interacti ons, including those that involve environmental hazards and disasters, (C) 1998 Elsevier Science Ltd, All rights reserved.