Interpretation of first-order recurrent neural networks by means of fuzzy rules

Abstract

First-order recurrent neural networks can be trained to recognize strings of a regular language. Finite state automata can be extracted from these neural networks. Normally, a search process in the output domain of the neurons is necessary for carrying out this extraction procedure. On the other hand, studies about fuzzy rules extraction from feedforward multilayered neural networks can be considered to define new techniques that transform first-order recurrent neural networks into finite state automata. With these new techniques, a fuzzy description of the action of each neuron can be obtained. From these descriptions, the transition function of the automaton can be directly found and, in this way, the search process is not necessary. A technique with this approach is presented in this paper. Besides, the used method to extract fuzzy rules from a neuron has the advantage that the inputs of the fuzzy system coincide with the inputs of the neuron. Thus, the fuzzy system is more intuitive. Once the transition function is obtained, the automaton structure can be found with the analysis of the transitions for every state and input from the initial state. Finally, several examples are presented to illustrate the method.