用于前馈神经网络的遗传设计

在人工神经网络应用中，由于存在网络规模和拓扑结构难以预先确定，网络学习速度慢，且易于收敛到局部最优点等问题，有关文献提出了采用基于遗传算法（ＧＡｓ） 思想进行设计和学习的方法，该方法能够同时确定网络的结构及有关参数。该文在此基础上，对此方法进行了改进，改进之处在于，采用浮点数矩阵来表示编码，同时对于遗传算法的进化过程也进行了一定的改进，使该方法能够接受一定的约束条件。针对前馈型神经网络，该方法在满足一定约束条件的情况下，能同时有效地寻找到合适的网络结构和相应的参数（ 神经网络的权值和阈值） ， 新方法较原方法在精度和速度上都有较大的提高。

A Genetic Design for Feedforward Neural Network

During application of neural network, there exist some problems,including difficult determination of the size and structure of neural network in advance, the learning speed of neural network is slow, and it's easy to converge to local optimum. In the view of these problems, some references proposed to use Genetic Algorithms (GAs) to design and train neural network, the top structure and related parameters (weights and thresholds) can be obtained simultaneously with this method. On this basis, this presentation made some improvement on proposed method. Main improvement includes that float point matrix is adopted to encode, evolution of GAs itself is modified and the proposed method can satisfy some constrain conditions. For feedforward neural network, this method can find suitable network structure and corresponding parameter (weights and thresholds) simultaneously under certain constrain condition. New method has great improvement over the old one in both accuracy and speed.