Searching neural network structures with L systems and genetic algorithms

Abstract

We present a new method for using genetic algorithms and L systems to grow up efficient neural network structures. Our L rules operate directly on 2-dimensional cell matrix. Lrules are produced automatically by genetic algorithm and they have“age”that controls the number of firing times, i.e., times we can apply each rule. We have modified the conventional neural network model so that it is easy to present the knowledge by birth (axon weights) and the learning by experience (dendrite weights). A connection is shown to exist between the axon weights and learning parameters used e.g., in back propagation. This system enables us to find special structures that are very fast for both to train and to operate comparing to conventional, layered methods.

Keywords:

• Genetic algorithms
• Lindenmayer systems
• back propagation
• network structure
• xor problem

C.R. Categories:

• I.2.6
• J.3
• F.4.2