论文部分内容阅读
为改善小波网络的学习和逼近性能 ,基于递阶结构染色体提出了递阶进化算法用以实现小波网络的设计和训练。该算法采用包含控制级基因和参数级基因的递阶结构的染色体 ,分别对网络结构和网络参数进行编码 ,并根据编码特点将遗传算法与进化规划结合进行进化操作 ,可实现同时对网络结构与网络参数进行进化设计和学习训练。该算法不仅克服了梯度下降算法中的局部极小和网络训练不收敛问题 ,也使网络结构更优 ,从而提高了网络训练效率和网络的工作性能。就函数逼近问题和水轮机组的状态预测问题进行了事例研究 ,验证了所提出的算法的优越性和可行性
In order to improve learning and approximation performance of wavelet networks, a hierarchical evolutionary algorithm based on hierarchical chromosomes is proposed to realize the design and training of wavelet networks. The algorithm uses a hierarchical structure of chromosomes containing control-level genes and parameter-level genes to encode the network structure and network parameters respectively. According to the coding characteristics, genetic algorithms and evolutionary programming are combined for evolutionary operations, Network parameters for evolutionary design and learning training. This algorithm not only overcomes the problem of local minima in gradient descent algorithm and the non-convergence of network training, but also improves the network structure and improves the network training efficiency and network performance. Case studies are made on the function approximation problem and the state prediction of hydro-turbine, which verifies the superiority and feasibility of the proposed algorithm