Abstract
Spatial Analysis (SA) techniques are important and becoming even more so as the supply of spatial data is increasing. The 1990s are witnessing the beginnings of a new paradigm for spatial analysis, based upon novel modes of computation which are collectively known as computational intelligence (CI) technologies. The driving force is a combination of large amounts of spatial data due to the GIS data revolution, the availability of attractive and novel CI-tools, the rapid growth in computational power, and the new emphasis on exploratory data analysis and modelling.
We understand the term computational intelligence in the sense as suggested by James Bezdek at the 1994 IEEE-World Congress on CI held at Orlando. It denotes the lowest level forms of intelligence which stem from the ability to exhibit some form of computational adaptivity and fault tolerance, without explicitly representing knowledge in the AI sense. Artificial life, evolutionary computation, intelligent agents and neural networks are major components in this arena which hold some promise to meet the new large scale data processing needs in GIS.
No doubt, CI is best designed in capturing those systems which can efficiently process information in a massively parallel way and learn by adjusting certain parameters. Thus, we limit the scope of our discussion essentially to neural networks (NN), briefly characterize the fundamentals of NN driven SA and exemplify its potential for two major SA tasks (spatial interaction modelling and pattern classification).
Neural networks - viewed as generalizations of conventional spatial data analysis techniques and models - provide spatial analysts with rich and flexible classes of novel non-linear data-driven mathematical tools. The application of NNs holds the potential for fundamental advances in empirical understanding across a broad spectrum of application fields in SA. To realize these advances, it is important to adopt a principled rather than ad hoc approach where spatial statistics and neural network modelling have to work together. The most important challenges in the years to come will be twofold: first, to develop sound application domain specific methodologies, and, second, to gain deeper theoretical insights into the complex relationship between learning and generalization which is crucial for the success of real world applications.