Adaptive Systems and Evolutionary Neural Networks : a Survey M.Annunziato 1 , M.Lucchetti 2 , S.Pizzuti 1,3 1 ENEA – Energy, New technologies and Environment Agency, ‘Casaccia’ R.C. Via Anguillarese 301, 00060 Rome, Italy Phone: +39-06-30484411, Fax: +39-06-30484811 email:{mauro.annunziato, stefano.pizzuti}@casaccia.enea.it 2 University of Rome ‘La Sapienza’ Department of Computer and Systems Science Via Eudossiana 18, 00184 Rome Italy Phone: +39-06-44585938, Fax: +39-06-44585367 email: lucchetti@dis.uniroma1.it 3 CS – Communication Systems S.p.A. Piazza della Repubblica 32, Milan Italy ABSTRACT: During last decades there has been an increasing interest in artificially combining evolution and learning, in order to pursue adaptivity and to increase efficiency of control, supervision and optimisation systems. In particular the need for adaptation came out from several real-world applications in non-stationary environments ranging from non linear control tasks to manufacturing process optimisation, from time series forecasting to interactive game playing. These needs led to the birth of a new general framework for adaptive systems, namely the Evolutionary Artificial Neural Networks, where the modelling potentialities of artificial neural networks have been matched with the adaptation properties of the evolutionary algorithms. This paper briefly reviews the main results achieved and presents the state-of-the-art in this field. KEYWORDS: evolutionary neural networks, adaptive systems, non-stationary environment, neural networks, evolutionary computation INTRODUCTION In several real-world dynamical systems applications, ranging from robotics to telecommunications, from process automation to biology simulations, it’s impossible to formulate an a priori exact model of the system, taking into account all the variables influencing the evolution during time, partly because of the presence of some unobservable dynamics in the system, which result in non-stationary phenomena on the measured data, and partly because of the unavoidable and unpredictable noise that affects the system internal state and its output. In such cases a possible approach is to extract a first rough offline model, and then to on line update it, in order to recover the gap of knowledge about unknown disturbances and to pursue an on line adaptation of the model itself. This task has been the research subject of many groups involved in the development of smart adaptive systems for real applications. One of the most commonly ascribed approach to reach adaptive continuous updating consisted in matching the neural networks modelling capabilities with the adaptation properties of evolutionary algorithms. These investigations led to the birth of a new framework which is generally referred to as Evolutionary Artificial Neural Networks (EANNs). The model we get from a neural network undergoes the evolution superimposed by the artificial environment of the evolutionary algorithm which is related to the evolution of the system. In such a way it’s possible to couple the evolution of the population and the learning process of each individual of the population, achieving better adaptation of the whole environment to a generic dynamic fitness landscape. EVOLVING NEURAL NETWORKS: THEORETICAL ISSUES AND PREVIOUS WORK The first attempts to conjugate evolutionary algorithms with neural networks dealt with the offline set-up of the networks (i.e. training of the connection weights [1], offline design of the neural architecture [2][3], or both [4][5]). This type of coupling is still implemented in many applications (see for example [6] [7] [8]). However the achieved