NEURAL NETWORKS DESIGN USING GA WITH PLEIOTROPY EFFECT +DOLQD.ZDQLFND6HEDVWLDQ5\QNLHZLF] 'HSDUWPHQWRI&RPSXWHU6FLHQFH:URFáDZ8QLYHUVLW\RI7HFKQRORJ\ :\E:\VSLDVNLHJR:URFáDZNZDVQLFND#FLSZUZURFSO Abstract The paper presents results of designing neural networks using evolutionary algorithm with new representation schema. The pleiotropy and polygenic effect is implemented in the scheme of individuals coding. The influence of this effect on evolutionary algorithm efficiency is the main subject of the simulation study. The results obtained using proposed algorithm as a tool of neural network design are compared with outcome of neural network design using classic genetic algorithms. 1. Introduction Neural networks (NNs) are strongly developed fields, started from the earlier forties when McCulloch and Pitts proposed their architecture. The similarities of artificial and biological neural networks lie on possibilities of learning on the base of training set, and generalizing. In the last decades new methods of training have been developed, as well as their mathematical basis. NNs are useful in variety of practical applications (Freeman, Skapura 1992). Genetic algorithms (GAs) are also a technique from the widely understood area of artificial intelligence, based on the natural process – biological evolution (Holland 1975, .ZDQLFND  7KH PDLQ GLIIHUHQFHV EHWZHHQ FRQYHQWLRQDO RSWLPL]DWLRQ PHWKRGV DQG genetic algorithm are: • GAs work with coded parameters in the form of chromosomes. Usually chromosomes are bits strings. One chromosome (individual) codes single point on the solutions space. • GAs search the solution working simultaneously with a population of individuals. • GAs do not use derivatives or any other information about optimized function. • GAs use probabilistic rules during search process, exploiting areas with high fitness. Recently, the hybrid systems becomes strongly developed and used to solve many practical tasks (Medsker 1995). Combination of different techniques can give better solution or the solution can be obtained in shorter time. In this paper we focus on the combination of GAs and NNs. However, various approaches to combination of GAs and NNs are observed, we focus on GAs used to NNs design. The main difference between proposed approach and others (Bornholds et al. 1992, Schafer et al. 1992, Kucsu, Tronton 1994) lies in the specific operators of used GA, namely – pleiotropy and polygenic effect (so called GAPP). 2. Designing neural networks An artificial neural network is the collection of connected units, called neurons. The connections are weighted and weights are usually real values. NN consists of inputs neurons