基于改进遗传算法的BP神经网络自适应优化设计概要

0.99 5

9第30卷第4期 柴 毅,等: 基于改进遗传算法的BP 神经网络自适应优化设计

96 重庆大学学报 (自然科学版 2007 年 ing Company Inc, 1990. [ 2 ] DJ MONTANA , L Da Training Feedforward Neural Networks U sing Genetic A lgorithm s [ C ] ∥ Proceedings 11 Morgan Kaufman, San M ateo CA: 1989. 762 2 767. th th 4 结束语 长期以来 , 由于对神经网络的拓扑结构设计和初 始权值的设置 , 缺乏理论支持 . 使得所设计的网络往往 存在网络规模庞大 , 设计效率低下与经典 BP 算法收 敛过慢 、 局部收敛等问题 . 这常常是神经网络得不到有 效应用的因素之一 . 本文使用遗传算法在全局解空间 内对网络拓扑结构和网络参数进行自适应的动态调 整 , 从而获得网络的最优设计 . 克服了由于神经网络初 始权值的随机性和网络结构确定过程中所带来的网络 振荡 , 以及网络极易陷入局部解问题并且有效提高神 经网络的泛化能力 , 这种方法避免了一般神经网络依 靠经验确定网络结构的困难 . 并将该方法应用到番茄 常见病害预测的实例中 , 收到了比较好结果 . 参考文献 [ 1 ] GOLDBERG D Genetic A lgorithm s in Search, Op tim ization and M achine Learn [M ]. London: Addison2 esley Publish2 W tional Joint Conference on A rtificial Intelligence. DetroitM I: [ 3 ] H KITANO. Em irical Studies on the Speed of Convergence of Neural Network Training using Genetic A lgorithm s [ C ] ∥ gence, 1990, Boston MASS, 1990: 789 2 796. th [ 4 ] H KITANO. Emp irical Studies on the Speed of Convergence of Neural Network Training using Genetic A lgorithm s[ C ] ∥ ligence. Boston MASS, 1990: 789 2 796. [5] 虞和济 . 基于神经网络的智能诊断 [M ]. 北京 : 冶金工业 出版社 , 2000. [6] 阎平凡 . 人工神经网络与模拟进化计算 [M ]. 北京 : 清华 大学出版社 , 2005. ( College of Autom ation, Chongqing University, Chongqing 400030, China Abstract: B P (Back Propagation Neural networks is in the p resence of the local op tim ization in the Neural networks training The algorithm have slow convergence and the local convergence p roblem which impact the

neural networks work . lem in the traditional NN designing, the action p rincip les of B P 2 Neural network ’s structure are analyzed, and a new speed. Tomatoes disease diagnosis examp les illustrate the feasibility of this app roach. (编辑 张小强 method is for med which is confir ed from the Enhance genetic algorithm s ( EGA . The method can identify net ork con2 m w figuration and net ork training methods B y adop ting the num ber coding, self2adap table multi2 w . point variations operation, Key words: i p roved genetic arithmetic; EGA; B P arithmetic; m ulti2layer sensor; NN Structure m this method can effectively reduce the network size and the net ork convergence ti e, increase the network training w m perfor ance. In order to cover these shortcom ings and solves the size’s hugeness and the low efficiency of the net p rob2 m Self2 adapta tion O pti ize BP Neura l Network m D esign Ba sed on the Genetic A lgor ithm s CHA I Yi, Y I Ho ng 2 e ng, L ID a 2jie , ZHAN G Ke N p Proceeding 8 JMJT National Conference in A rificial Intelli2 Proccedings 8 JM IT National Conference in A rtificial Intel2 Interna2