基于matlab的神经Mealy机的构造

【摘要】  递归神经网络相较于网络有着较强的逼近能力，一定的学习能力；Mealy机是现代计算机的一类，它由电路确定，运行时无法避免灾难性的后果。

本文构造得到的新型神经Mealy机，主要是通过学习来获得计算机结构，可以较好地避免。其本质是将递归神经网络通过BP优化算法，对Mealy机进行模拟得到。

基于形式文法和自动机的等价性，用神经网络来实现文法推导。先采用神经网络对样本集进行学习，这些样本可由一个经典Mealy机生成，然后从训练完的神经网络提取出自动机。主要是通过递归神经网络的BP学习算法逐渐调整各神经层的权值和阈值，使得误差最小化，即实际输出和期望输出一致。然后对神经Mealy机的性能做进一步提升，让它学习复杂的并有实际意义的Mealy机，制造出算法级别的神经计算机。最后实验结果表明，神经Mealy机的性能优于经典Mealy机，这说明神经Mealy机的构造是有意义的。

【关键词】  递归神经网络，Mealy机，BP算法，Matlab软件编程

The structure of the neural Mealy machine based on Matlab

【Abstract】   Compared with other networks，recurrent neural networks has strong ability of approximation and good learning ability; Mealy machine is a kind of modern computer, it was determined by the circuit, and the runtime cannot avoid catastrophic consequences.

In this paper, we construct the new type of neural Mealy machines, mainly for computer architecture by learning, can be avoided the catastrophic consequences. Essentially, with BP optimizing algorithm of recurrent neural network, it can be obtained after simulating Mealy machines.

Based on the equivalence of formal grammar and automata, we can use neural network to realize grammar derivation. First neural network learns the sample that may be produced by a classical Mealy machine, and then we get an automata from training the neural network. Mainly through the recurrent neural network BP learning algorithm gradually adjust the neural layer weights and thresholds, minimize the error, namely, make the actual output and the correct output consistent. Then we improve the performance of neural Mealy machine, let it learn complex and practical Mealy machines, we produce neural computer on algorithm level. At last, the experimental results show that the performance of neural Mealy machine is superior to the classic Mealy machine, hence it is meaningful to construct neural Mealy machines.

【Key Words】   recurrent neural networks，Mealy machine，BP algorithm，Matlab software programming calculation