Roberts, Ian and Hunter, Andrew (1999) Evolving neural networks using matrix grammars. Project Report. University of Sunderland, Sunderland.
Documents 

PDF
Evolving_Neural_Networks_Using_Matrix_Grammars__AH.pdf  Whole Document Restricted to Repository staff only 3MB 
Item Type:  Paper or Report (Project Report) 

Item Status:  Live Archive 
Abstract
Methods of evolving Neural Networks using Matrix Grammars are described. Because these methods generate network architectures structurally, reusing symbols to describe subsections of the architecture, they tend to produce wellstructured networks and are suitable for similarly wellstructured problems. Methods which generate the architecture only, and methods which also generate weights, are described. Evolution is combined with backpropagation training. The techniques are compared with previously published work, and show several distinct advantages. The main advantage of all the methods is the ability to overcome the Genetic Algorithm scaling problem. The inclusion of weights gives better convergence. The Matrix Grammars presented here further separate the evolution of weights and architecture than previous methods, widening the searchspace. The suitability of the techniques for more substantial problems is discussed. We also show how large improvements can be achieved by progressive evolution: the pretraining of the population on related, simpler problems.
Additional Information:  Methods of evolving Neural Networks using Matrix Grammars are described. Because these methods generate network architectures structurally, reusing symbols to describe subsections of the architecture, they tend to produce wellstructured networks and are suitable for similarly wellstructured problems. Methods which generate the architecture only, and methods which also generate weights, are described. Evolution is combined with backpropagation training. The techniques are compared with previously published work, and show several distinct advantages. The main advantage of all the methods is the ability to overcome the Genetic Algorithm scaling problem. The inclusion of weights gives better convergence. The Matrix Grammars presented here further separate the evolution of weights and architecture than previous methods, widening the searchspace. The suitability of the techniques for more substantial problems is discussed. We also show how large improvements can be achieved by progressive evolution: the pretraining of the population on related, simpler problems. 

Keywords:  Neural networks, Matrix grammars, Network architectures 
Subjects:  G Mathematical and Computer Sciences > G400 Computer Science 
Divisions:  College of Science > School of Computer Science 
ID Code:  3387 
Deposited On:  26 Sep 2010 16:44 
Repository Staff Only: item control page