Automaton Learning Network Theory

Abstract

Universal Learning Network (ULN) which can be used to model and control large scale complicated systems has been reported. The ULN is made up of nonlinearly operated nodes and multi-branches with arbitrary time delays between the nodes. Because of that, the ULN is capable of modeling in a natural way the large scale complicated systems which are hard to be modeled by the commonly used neural networks (NN). But, the ULN can't be applied to discrete event systems, because the ULN can only treat the system whose behavior is continuous.
In this paper, a discrete event oriented learning network which is called Automaton Learning Network (ALN) is presented, which is also based on the architecture of the ULN. As an ALN is a discrete event system, leaning algorithm such as Back Propagation in NN, which is based on differential calculus can not be utilized for the ALN. Therefore a new local search optimization method which has the capability of intensification and diversification is proposed.This method is called RasVan which is an abbreviation of Random Search with Variable Neighborhood. One of the features of RasVan is that when there is quite a possibility of finding good solutions around the current one, intensified search for the vicinity of the current solution is carried out, on the other hand, when there is no possibility of finding good solutions, diversified search is executed in order to find good solutions in the region far from the current solution. Finally, simulations of a simple ALN which is composed of a controlled automaton and a control automaton are carried out in order to study the dependence on the initial states in the ALN and study the fundamental characteristics of RasVan.