论文部分内容阅读
为了解决一类径向基函数神经网络的结构优化问题，提出了一种有效的混合优化策略。将结构优化问题转化为一类组合优化问题。利用结合遗传算法群体并行搜索能力和模拟退火概率突跳特性来改善优化效率并避免局部极小的混合策略。借助于有效的编码方式在结构解空间中优选网络结构。利用梯度下降计算隐层到输出层的连接权。增添和删除操作用于增加结构搜索的灵活性。为保证所得网络具有较好的推广能力，利用训练误差和检验误差的综合指标作为算法择取优良网络的依据。仿真研究表明，所提混合策略是快速有效的，且能保证网络具有较好的推广性和抗噪声能力。
In order to solve the structural optimization problem of a class of radial basis function neural networks, an effective hybrid optimization strategy is proposed. The structural optimization problem is transformed into a combinatorial optimization problem. Combining the ability of parallel searching of population of genetic algorithms and the sudden jump probability of simulated annealing, the optimization efficiency is improved and the local minima mixing strategy is avoided. The network structure is optimized in the structure solution space by means of efficient coding. Use gradient descent to calculate the connection of hidden layer to output layer. Add and delete operations are used to increase the flexibility of structure search. In order to ensure that the resulting network has a good promotion ability, the comprehensive index of training error and test error is used as the basis for choosing an excellent network. Simulation studies show that the proposed hybrid strategy is fast and effective, and can guarantee the network has better promotion and anti-noise ability.