考虑问题随机性的神经网络.PDFVIP

16 5 系 统 管 理 学 报 V ol. 16 N o. 5 2007 10 J ournal o f System s M anagement O ct . 2007 : 1005- 2542 2007) 05- 0563- 05 1 1 2 潘 华 , 李本强 , 雷元新 1. , 529020; 2. , 52 8000) = 考虑问题随机性的神经网络研究是一个 有挑战性的问题它涉及网络训练误差允许值的选 取以及网络预测结果的评价目前网络训练误差允许值的选取采用试算法, 而网络预测结果常常采用相对 误差法来评价, 认为相对误差越小预测精度越高对于有随机因素影响的样本数据, 已有方法不再适用因 此, 结合混凝土配合比设计问题, 考虑训练样本的随机特性, 研究了一种新的选取训练误差初始值方法和预 测结果评价方法, 并提出了适合随机问题的计算式最后进行了实例计算分析, 结果表明, 该方法可行, 评价 结论更贴近工程实际 : 神经网络; 预测; 随机 : T U 4 : Neural Network Study on Random Problem 1 1 2 PA N H ua , L I B en-q iang , L EI Yuan-x in 1. Wuyi Univ. , Jiang men 529020, China; 2 . F oshan U niv. , F oshan 52 8000 , China) =Abstract T he st udy o n art ificial neural net w ork dealing w it h random sam ples is a challenging pr oblem. For an artif icial neural net w ork mo del, t he choice of t he err or o f tr aining and t he ev aluat ion of t he predic- t ed result s is concerned . T o our know ledge, t he erro r of t raining is mo st ly chosen by t ria-l and- error . T he predict ed result s w ere of ten evaluat ed by u sing r elat ive error method , and it is o ft en reg arded t hat the less the r elat ive error is, t he bet ter t he predicting precision is. But t hese met hods are no longer f itt ed for ran- dom t raining sam ples. T herefo re, combined w ith the design of concret e mix proport ion, by taking int o acco unt t he st ochastic characterist ic of t raining samples, a new m et hod f or det erm ining t raining er ror and ev aluat ing predict ed r esult s of a neural net w ork m odel w as s