Self-Organizing Evolution Strategies

Self-organizing evolution strategies present a new combination of artificial neural networks and evolution strategies. While evolutionary algorithms were used for training or optimization of neural networks, the new combination reverts this principle by applying concepts of neural networks to evolution strategies. Especially, concepts of self-organizing maps (SOM) are to be applied to the evolution strategies.

The main characteristics of a SOM are the neighborhood relationship defined between its neurons and that the best neuron pulls its neighbors towards itself. Applying these characteristics to self-organizing evolution strategies, the individuals - which in common evolution strategies are not related to each other - are now arranged in a neighborhood relationship and the individuals with a higher fitness attract their less fit neighbors by performing a kind of "directed mutation". Thus, the individuals initially spread over the problem space are expected to concentrate around the optimum.

A simulation software is being developed to investigate the practical relevance of the self-organizing evolution strategies. So far, the individuals are arranged in an elastic grid. The initialization tries to spread the individuals to cover the entire problem space.

Illustration 1 depicts the distribution of individuals in a simple 2-dimensional example, after their initialization, the 3rd generation and the 15th generation.



Illustration 1: Contraction of the individuals-grid around the optimum in point (0;0)

It is planned to integrate this program into EvA and to conduct further refinements.

The main focus of this research is on areas of application for self-organizing evolution strategies and on putting them into relation with other optimization strategies.