MUTATION INFLUENCE AND PROBABILITY IN NEURO-GENETIC EVOLUTION AND NEAT. STRATEGY FOR ADJUSTING DYNAMIC MUTATION PARAMETERS IN RESPONSE TO FITNESS STAGNATION

Abstract

Optimization tasks are one of the hot nowadays challenges in science and techniques which can be resolved efficiently by evolutionary algorithms. Although they have proved their efficiency in neuroevolution they have to compete with other methods of neural network training. They began to show their advantages with the development of computing technologies and proved themselves well with problems where more traditional methods, such as gradient descent, cannot be applied. It would be great to investigate already existing neuroevolutionary methods or develop new ones.

A different configuration of training parameters can have a significant impact on the result. Thus the investigation in this direction is actual. Inspecting the influence of training parameters and various strategies for tuning these parameters we found interesting and actual the challenge of automatic adjustment of training parameters during training.

The main goal of the paper is an investigation of the training parameter’s influences on the training results of two popular neuroevolution methods: the neuro-genetic evolution with a static topology of the neural network and the neuroevolution method of augmenting topologies - NEAT.  A mutation step (mutation strength) is a studied parameter for the first method, and for the NEAT we have two parameters: a mutation step and a probability of structural mutations. The proposed strategy for the automatic adjustment of the studied parameters aims to solve the problem of reaching a local extremum, which didn’t allow to achieve the best possible results of training.

All experiments were performed on the same task of a rocket controlling the flight from the Earth to the Moon, which requires enhanced accuracy of the neural network for successful training. The results of the training were analyzed according to the following indicators: average fitness value in the generation, best fitness value in the generation, and percentage of successful individuals in the generation. The conducted experiments have demonstrated that the requirements with high accuracy force usage of lower values of the mutation strength for traditional genetic algorithms, but this creates risks of getting stuck in local extrema or convergence to a suboptimal solution.

The NEAT experiments have shown that the task’s requirements for high accuracy didn’t allow the best results to be achieved. The proposed strategy for decreasing the mutation strength and the probability of mutations was able to reduce the risk of convergence to a suboptimal solution in both methods. Generally, this approach was able to justify itself. Further investigations of this strategy on par with already existing approaches might reveal other advantages which should be investigated more deeply.

Keywords: genetic algorithms, neuroevolution, artificial neural networks, NEAT