A Model of Recurrent Networks that Learn the Finite Automaton from Given Input-Output Sequences

From the viewpoint of applying recurrent neural networks to AI, it is not suitable to use learning algorithms based on optimum control theory. One reason of this problem is lack of correspondence between these algorithms and traditional symbol processing. In this report, we propose a new algorithm for modi ed Elman networks (PEX model). This algorithm is derived from the minimization procedure of nite automata under the correspondence between Elman networks and nite automata. In this algorithm, a network predicts next network state by itself, and this prediction plays an important role in re ning the network. In addition, we examined behavior of this model toward noise and found the network gets robust state-transition against noise. Moreover, we discuss the possibility to deal with CFG by an experiment of sub-grammar learning.