Learning regular languages via recurrent higher-order neural networks
Ken Tanaka, Itsuo Kumazawa · 2002
Learning regular languages is accomplished by the acquisition of finite state automata. In order for a neural network to acquire an arbitrary FSA, the network must first have a representation for every FSA state, and furthermore be able to realize an arbitrary FSA state transfer function. We show that if the network model of Giles et al. (1992) represents each FSA state using local state representation, then it can realize any FSA state transfer function. However, this may be difficult to acquire by learning, and is a reason why the model is not necessarily successful at learning some regular languages. In order to overcome this problem we propose the recurrent higher-order neural network (RHON). We show the order of connections sufficient to realize any FSA state transfer function regardless of the networks representation of states. After deriving the learning algorithm, we show the learning superiority of RHON to the model of Giles et al. through computer simulation.