2006 年 19 巻 2 号 p. 51-58
A novel approach to Topology and Weight Evolving Artificial Neural Networks (TWEANN) is presented. This comes from the following two considerations. (1) Artificial evolution is designed without recombination which mostly generates an offspring whose fitness value is considerably reduced than those of parents. Instead of recombination, two types of topological mutations are provided as genetic operations. One is for inserting a new neuron and the other is for generating a new synaptic connection. The both of them are designed carefully to retain the current evaluation value. (2) A new encoding technique such that a string is defined as a set of substrings called operons is introduced. This is based on our idea that each potential function should be encoded into a different part of genetic information. In order to validate our approach, computer simulations are conducted by using one of the standard reinforcement learning problems called the double pole balancing problems without velocity information. The results obtained are compared with the results by NEAT, which is recognized as one of the most powerful techniques in this problem domain. It is found that our proposed approach yields competitive results, especially when the problem is tuned to be difficult.