Inference of stochastic regular languages through simple recurrent networks

Grammatical inference has been recently approached through artificial neural networks. Recurrent connectionist architectures were trained to accept or reject strings belonging to a number of specific regular languages, or to predict the possible successor(s) for each character in the string. On the other hand, for static (non-string) data, M.D. Richard et al. (1991), showed that a nonrecurrent architecture can estimate Bayesian a posteriori probabilities. The authors show empirical evidence supporting this statement which also seems to be verified when simple recurrent networks (SRNs) are used to estimate probabilities of stochastic regular languages