摘要： 针对神经网络结构与参数并行优化问题，提出一种基于动态多群体差分进化算法的前向神经网络设计方法。采用分层递阶结构原理构造算法个体，根据控制基因信息将个体分成不同的动态群体。通过对个体进行重构，实现进化过程中个体信息的充分交换与共享。设计基于群体适应度的控制基因更新方法来优化网络拓扑结构，克服结构优化的盲目与低效问题。将所设计的神经网络应用于大包线飞行控制律参数拟合中。仿真结果表明，该算法能快速有效地确定神经网络的结构和权值，所优化的网络在调参控制中具有较好的泛化能力。

关键词: 神经网络, 结构优化, 动态多群体, 个体重构, 差分进化算法, 增益调参

Abstract: This paper proposes a method of designing feedforward neural network based on Dynamic Multi-group Differential Evolution(DMDE) algorithm, which is introduced into hierarchical structure theory, and divides population into multi-group individuals dynamically. The new algorithm reconstructs individual gene to achieve information sharing and exchanging between individuals. To overcome the problem of optimizing structure blindly and inefficiently, the control gene is updated by group fitness. Simulation suggests that DMDE algorithm can optimize the topology structure and weights of the feedforward neural network efficiently. Furthermore, the designed neural network has high generalization ability when applied in gain schedule control.

Key words: neural network, structure optimization, dynamic multi-group, individual reconstruction, Differential Evolution(DE) algorithm, gain scheduling

中图分类号: 

• TP393