2. Neural Networks
With the development of NN, which are Artificial Intelligence based technologies, in
recent years news opportunities have emerged to enhance the tools we use to process
spatial Data. Their specific advantage relies non only in the enhancement of speed and
efficiency in handling urban Data, but specifically in providing a tool to develop new
theories and techniques. While the traditional modelling approach is based on explicit
“a priori” rules formulation, through an AI connessionistic approach rules are found “a
posteriori” on the base of a learning process of a distributed “unit processing”
architecture.
NN model is a parallel distributed Information system consisting of a set of adaptive
processing elements (nodes) and a set of unidirectional data connections (weights).
The most successful applications in territorial Analysis and Planning rely on pattern
classification, clustering or categorisation, optimisation (Openshaw and Abrahart 2000;
Reggiani 2000; Leung and Fischer 2001), modelling scenic beauty from extracted
landscape attributes (Bishop 1994), suitability analysis for development (Sui 1992;
Deadman and Gimblett, 1995).
The novelty of our approach lies in the use of NN as a powerful tool for prediction and
building virtual scenarios on urbanisation process. The results have been achieved
through different categories of “training regimen” able to react to different information
environment.
The training processes can be divided into three basic categories: monitored training,
supervised training, and self-organisation. The monitored training is typical of
associative networks, which are NN with essentially a single functional layer that
associated one set of vector Xi, x2, ...xn with another set of vector yι, y2, ...yn. The
primary classification of ANN are into feedforward and recurrent classes. Another
categorisation of ANN is into autoassociative NN if y vectors are assumed to be equal
to the corresponding x vectors. In a Heteroassociative network yi ≠ xi.
There are many algorithms and procedures to optimize the weight matrix during the
learning phase and many algorithms for dynamically query the ANN already trained.
In this research we used a Recirculation Neural Network (RCNN) (Hinton and
McLelland, i988). The ANN have shown to be highly efficient in determining the fuzzy
similarities among different Records in any Data Base (DB) and the relationships of
gradual solidarity and gradual incompatibility among the different Variables. The ability
of ANN to produce prototypical generators, to discover ethnotypologies and to simulate