Study, Based On Neural Networks, Fuzzy Finite State Automata

Fuzzy Gramma has become the major developing fields of the research of fuzzy finite state automaton because of the equivalence. More over, fuzzy grammatical inference has extensive application and distinct advantage in Syntactic recognition of skeletal maturity form X-rays, Detecting and quantifying fuzzy artery lesions from arteriograms, Intelligent interface design, Clinical monitoring, Lexical analysis, etc. Furthermore, grammatical inference is a suitable technique for solving the problems such as speech recognition, object recognition in images, protein structure prediction, gene structure prediction, etc. In addition, the fuzzy regular grammar is the simplest and most widely used. Therefore, the research of fuzzy regular grammar inference based on neural networks has great practical significance and high theoretical value.In this thesis, we introduce the basic theory of fuzzy automaton begin with Chomsky hierarchy. And then, we introduce the application of the neural networks approach in fuzzy finite state automaton. It mainly includes a complete system of fuzzy automaton inductiong using neural network, concretely , the relationship of fuzzy finite state automaton and neural networks and fuzzy gramma, the training of neural networks based on sample set of fuzzy finite state automaton, the fuzzy grammatical inference used neural networks, the extraction algortthims of fuzzy finite state automaton. In this thesis, our ground work centers on the algorithms of fuzzy grammatical inference based on the neural network. There are two main algorithms available real-time recurrent learning (RTRL) algorithm and real-coded genetic algorithm (RCGA). But, the rate of convergence of the both methods in the fuzzy grammatical inference used neural networks is rather slow, the time complexity could be rather high, and weak generative ability. Besides, RTRL algorithm is unstabkle and the RCGA algorithm could be trapped in prematurellty easily. Worse, both of them and the abnormal conditions occurred frequently because of the data sources obtained in the experiment.According to the above-mentioned problems, we have done three works, and may be briefly summed up as follows:Based on the LMGA algorithm of nural network for fuzzy regular grammatical Inference to be raised. For improving the traditional RCGA algorithm, Levenberg-Marquardt algorithm is introduced into the process of optimal Selection, it not only improves the rate of convergence but also Skips the premature genes automatically. Expriment demonstrated that LMGA algorithm resolves the drawback of ill-conditioned prematurelity and improves the rate of maturation. The expriment also illustrates that LMGA algorithm can solve out the problems of the high complexity, week generative ablity and unstablity of RTRL algorithm..Based on the LMBP algorithm of nural network for fuzzy regular grammatical Inference to be raised. The Levenberg-Marquard(tLMBP) which is the fast algorthm of neural networks at present is introduced into the fuzzy regular grammatical inference. In this thesis, we discribe the algorithm and analysis the advantage and drawback by experiment. It is obvious that the convenge rate is high to perform the FRGI. It is also has the ability of processing super-long strings, too.Based on the VLBP algorithm of nural network for fuzzy regular grammatical Inference to be raised. VLBP algorithm is a Back-propagation algorithm that makes full use of the gradient information; it is the combinative product of heuristics information technique and RTRL algorithm.We use the methods such as comparison, induction, analysis, synthesis, etc for logical proof, and we do experiments to simulate and give examples to verify the feasibility of our algorithm.