The University of Edinburgh -
Division of Informatics
Forrest Hill & 80 South Bridge

MSc Thesis #9586

Title:Generating Neural Networks with Genetic Algorithms Using a Marker Based Encoding
Date: 1995
Abstract:Genetic algorithms and other evolutionary approaches are commonly used to generate neural networks. Marker based encoding is a novel approach to direct linear encoding of a neural network onto a chromosome representation for use by a genetic algorithm. Previous work has generated promising results using marker based encoding; its ability to evolve network structure as well as weights being a strong point, as this allows problems for which an appropriate network structure is not known to be tackled. We set out to examine how the encoding works, how sensitive it is to input parameters, find guidelines for appropriate setting of parameters, and how effective it is when compared to other methods and when scaled up to solve larger problems. It is concluded that the algorithm as it stands works more as a method of stochastic hill-climbing than a true genetic algorithm; it is nonetheless effective for certain types of problem, but more likely to be susceptible to entrapment by local peaks in the evaluation function surface than other genetic algorithms. Crossover is ineffective on a diverse population due to the competing conventions problem, but is successful at recombining mutations once convergence is close. Crossover has several side effects under MBE which are analysed. The algorithm is found to be fairly resistant to its parameter values, and guidelines are given for setting them effectively. It is unlikely to scale up well for large problems, but provides ideas and lessons for other evolutionary methods to draw on.

[Search These Pages] [DAI Home Page] [Comment]