摘要
多物种并行进化遗传算法应用于神经网络拓扑结构的设计,开辟了新的研究领域,论文提出伪并行(PPGA-MBP)混合遗传算法,结合改进的BP算法优化多层前馈神经网络的拓扑结构。算法采用实数编码来克服传统二进制编码的精度不足问题,并设计基于层次的杂交算子允许结构相异的个体杂交重组成新的个体,适应度函数更是综合考虑了均方误差、网络结构复杂度和网络的泛化能力等因素。实验证明取得了明显的优化效果,提高了神经网络的自适应能力和泛化能力,具有全局快速收敛的性能。论文还运用该算法建立了工业增产值经济预测网络模型,将网络预测值和多项式拟合值进行了对比分析。
Muhigroup parallel genetic algorithm has been introduced as a learning method form Muhilayer Feedforword Neural Networks（MFNN）.A novel approach,combined Pseudo-parallelism evolution technique based on sub-population competition with improved BP mechanism（PPGA-MBP）,is presented to evolve the weights and biases of all the MFNN.First,real encoding is introduced to solve the accuracy insufficiency problem of the traditional binary encoding.In addition,a new based-layers crossover operator is devised to enhance the algorithm performance,which allows that two net works with different number of units can be crossed to a new valid ＂child＂ network.Furthermore,fitness function is composed of mean squared error,network structure complexity and generalization ability.The experimental results show that it can solve the N-bit parity problem and enable to get the real-time information of population diversity during the process of evolution and has some improvements in both global converging velocity and searching precision.In the end,the best network is chosen as a solution to the economic forecasting problem.
出处
《计算机工程与应用》
CSCD
北大核心
2006年第13期73-76,共4页
Computer Engineering and Applications
基金
济南大学科技基金资助项目(编号:Y0425)
关键词
神经网络
伪并行遗传算法
遗传算法
经济预测
neural network,pseudo-parallelism genetic algorithm,genetic algorithm,economic forecast
