Work

Genetic neural networks on MIMD computers

Year: 1992
Type: dissertation
Abstract: This work draws together two natural metaphors: draw inspiration from natural evolution to attempt to provide a robust, efficient search technique, and networks form a crude model of information processing in brains which offer the prospect of computers to solve tasks by example. It is very natural to think of applying the adaptive techniques of genetic algorithms to the problem of searching the space of neural networks, but doing so is extremely hard, and provides the motivation for this work. It is argued that the key determinant of the success of any particular genetic algorithm is the interaction between the underlying correlations in the search space, the representation of the space adopted, and the genetic operators used to manipulate the representatives of elements in the search space. It is further argued that genetic algorithms as usually formulated are not ideally suited to training neural networks, and that in order to make significant progress in this area a broadening of the standard schema analysis is required. Such a generalisation, based on the notion of imposing suitable nested sets of equivalence relations over arbitrary search spaces, is proposed and developed. Design principles to help construct genetic algorithms for arbitrary problems are suggested in the context of such knowledge of the regularities in the search space as are known. The techniques developed are applied to a number of problem domains and yield some new insights. Issues of linkage and convergence are also relevant to the application of adaptive genetic techniques to neural network problems. Studies of these are presented. Existing attempts to apply genetic algorithms are also reviewed in the light of the non-standard analysis developed, and the prospects for further progress are discussed. In recognition of the fact that much of this work was carried Out on a large, medium-grained, reconfigurable parallel computer, a study of connection strategies for such machines is also presented.
Source: ERA
Language: English
Open Access status: green
OpenAlex ID: W1501587338
Cites: 41
Cited by: 56
Related to: 10