Evaluating the Learning of Automata through the Use of Recurrent Neural Networks

Abstract

This paper presents the computational potentialities, demonstrating the learning capacity and extrapolation of recurrent neural networks (RNN) through the learning study of finite automata. Initially, the generalized delta rule is presented for a recurrent network. In addition, the theoretical basis on Finite Automata, Moore Machine and Mealy Machine are presented, showing the equivalence between them. We formalized the problem of parity using a Moore Machine and its equivalent Mealy Machine. We proceeded simulation and the results are presented for the problem and  simulation of Tomita languages 1 and 2, testing the extrapolation capacity of each RNN. The results indicated that the neural networks with recurrent architectures have a good capacity of extrapolation for the simulated problems.