微粒群优化动态神经网络模型结构分析  

Model-structure analysis of dynamic neural networks with particle-swarm optimization

作　　者：范剑超[1] 韩敏[1] 

机构地区：[1]大连理工大学电子信息与电气工程学部,辽宁大连116023

出　　处：《控制理论与应用》2011年第9期1075-1081,共7页Control Theory & Applications

基　　金：国家自然科学基金资助项目(61074096);国家高技术研究发展计划"863"资助项目(2007AA04Z158);国家科技支撑计划资助项目(2006BAB14B05);国家重点基础研究发展计划"973"资助项目(2006CB403405)

摘　　要：微粒群算法由于进化机制中的随机不确定性,其稳定性很难进行分析,所以对微粒群的研究多是根据经验的实际优化模型求解.针对该问题,利用鲁棒不确定性理论,将算法分解为时不变和不确定时变的结构,减少原有参数固定的假设条件,从而对引入动态惯性权重的微粒群算法的渐近稳定性进行分析.在此基础之上,采用李雅普诺夫方法,得到基于微粒群参数优化的动态神经网络收敛的充分条件,自适应调整微粒速度的上下限,为组合模型的实际应用提供参数选择的理论基础.最后,通过仿真实例验证了所给出微粒群算法稳定性条件和基于微粒群优化的动态神经网络收敛条件的有效性.Due to the random uncertainty in the particle-swarm optimization algorithm（PSO）,the analysis of stability is difficult to be performed.Most of the researchers on PSO solve this problem based on the practical model obtained by experience.Being different,we employ the robust uncertainty theory to decompose the original algorithm into the time-invariant part and the uncertain time-variant part for reducing the original fixed constraints on parameters,and perform the asymptotic stability analysis by using the PSO algorithm with dynamic inertia weight.By using the Lyapunov method,we obtain the sufficient conditions of stability for the dynamic neural networks based on PSO and the upper and lower bounds of the parameters to be adjusted,providing the theoretical basis for parameter selection.Finally,simulation examples validate the stability conditions and the effectiveness of the proposed dynamic neural networks based on PSO algorithm.

关 键 词：微粒群优化 动态神经网络 鲁棒不确定性 稳定性 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]