A Neuro-Genetic Approach to Neural Network Design Antonia Azzini, Massimo Lazzaroni, and Andrea G.B. Tettamanzi Universit´a degli Studi di Milano Dipartimento di Tecnologie dell’Informazione via Bramante, 65 26013, Crema, Italy azzini,lazzaroni,tettamanzi@dti.unimi.it Abstract. This paper presents an approach to the joint optimization of neural network structure and weights which can take advantage of BP as a specialized decoder. The approach is validated on the toy problem of N-Input Parity Function and successfully applied to a real-world engine fault diagnosis problem. 1 Introduction Neuro-genetic systems have become a very important topic of study in recent years. Like indicated by Yao et al. in [14], they are biologycally-inspired com- putational models that use evolutionary algorithms (EAs) in conjunction with neural networks (NNs) to solve problems. The evolutionary approach is a more integrated way of designing ANNs since it allows all aspects of NN design to be taken into account at once and does not require expert knowledge of the problem. Much research has been undertaken on the combination of EAs and NNs. Through the use of EAs, the problem of designing a NN is regarded as an opti- mization problem. EAs can be applied to design ANNs in several ways. Some schemes concen- trate just on weight optimization: these can be regarded as alternative training algorithms; others have implemented a search over the topology space, or a search for the optimal learning parameters. The evolution of weights assumes that the architecture of the network must be static. The primary motivation for using evolutionary techniques to establish the weighting values rather than traditional gradient descent techniques such as BP [8], lies in the trapping in local minima and in the non-differentiability of the function. For this reason, rather than adapting weights based on local improve- ment only, EAs evolve weights based on the whole network fitness. Several works in this direction have been done by Montana and Davis [7] and by Whitley and colleagues [11]; in [12], they also implemented a purely evolutionary approach using binary codings of weights. In other cases an EA and a gradient descent algorithm have been combined [4].