一种结构自适应神经网络及其训练方法.pdfVIP

第25 卷 第8 期 控 制 与 决 策 2010 年 8 月 Vol. 25 No. 8 Control and Decision Aug. 2010 文章编号: 1001-0920 (2010) 08-1265-04 一种结构自适应神经网络及其训练方法 1 1,2 宋彦坡 , 彭小奇 (1. 中南大学能源科学与工程学院，长沙410083 ；2. 湖南第一师范学院信息科学与工程系，长沙410205) 摘 要: 针对神经网络建模效果对网络结构、训练方法过于敏感的缺陷, 提出一种结构自适应神经网络模型及其训 练方法. 模型具有双网结构并以“提前终止法”训练, 一定程度上降低了建模效果对网络结构的敏感性; 模型结构根 据建模数据的噪声方差、模型当前误差等信息自适应调整, 进一步提高了模型的建模效果, 同时具有较高的时间效 率. 仿真结果表明, 该方法弥补了提前终止等传统方法的部分不足, 具有较好的效果. 关键词: 优化建模；训练策略；神经网络；结构自适应 中图分类号: TP183 文献标识码: A New structure adapting neural network and its training method 1 1,2 SONG Yan-po , PENG Xiao-qi (1. School of Energy Science and Engineering，Central South University，Changsha 410083，China ；2. Department of Information Science and Engineering，Hu’nan First Normal University ，Changsha 410205，China ．Correspondent: SONG Yan-po，E-mail ：songyanpo@126. com) Abstract: The performance of artificial neural network (ANN) is very sensitive to its structure and training method. Therefore, an adaptive structure ANN and its training method are proposed. The ANN model is with double-net structure and trained by early stopping, which reduces the sensitivity of model performance to structure. The ANN structure is proposed to be adjusted dynamically according to the information of data noise and model error, which further improves the model performance with relatively high time efficiency. Simulation results show that the proposed methods can make up some defects of traditional methods such as early stopping, and has better performance. Key words: Optimal modeling ；Training strategy ；Neural network ；Structure adapting 1 引引引 言言言