摘要
人工神经网络的结构设计没有系统的规律可循,而基于梯度的神经网络参数优化又易于陷入局部最优解。该文研究了用带退化的协同进化遗传算法来优化神经网络结构,同时优化网络参数。将网络参数作为实数编码基因进行遗传选择,参数个体的受损率超过退化阈值时发生结构退化。退化进程由协同进化的控制个体动态控制。实验证明,该方案能够有效简化神经网络的结构和得到最优网络参数,收敛速度比常规遗传算法快。
Structural designing of artificial neural network is always a trouble problem without systematic rule and local minimum usually connects with conventional grads based on parameters optimization. Optimization of neural networks structure and parameters based on co-evolutionary genetic algorithm with degeneration are studied. Parameters are coded as gene co-evolving with control gene that controls degeneration. Experiments show that this approach get simpler structure, better parameters and more rapid convergence.
出处
《计算机工程与设计》
CSCD
北大核心
2007年第3期638-641,共4页
Computer Engineering and Design
关键词
遗传算法
协同进化
退化
神经网络
优化
genetic algorithm
co-evolution
degeneration
neural networks
optimization