用神經(jīng)網(wǎng)絡(luò)解決異或問題的新算法簡(jiǎn)介

頁(yè)數(shù) 18 字?jǐn)?shù) 5785

【摘要】從感知機(jī)的結(jié)構(gòu)及學(xué)習(xí)規(guī)則無法執(zhí)行“異或問題”出發(fā)，用神經(jīng)網(wǎng)絡(luò)中的BP網(wǎng)絡(luò)來解決“異或問題”,消除了感知器的局限性。但BP算法在具體實(shí)現(xiàn)中常會(huì)出現(xiàn)一些問題，如：收斂速度緩慢且與其它參數(shù)存在較強(qiáng)的耦合關(guān)系，局部極小等。對(duì)此，本作品從前饋神經(jīng)網(wǎng)絡(luò)的原理出發(fā)，提出一種變學(xué)習(xí)強(qiáng)度的速率適應(yīng)因子方法，用于對(duì)BP算法的改進(jìn)，并將改進(jìn)的算法用于二維XOR問題及多維XOR問題的學(xué)習(xí)。實(shí)驗(yàn)仿真證明，改進(jìn)后的算法可顯著加速網(wǎng)絡(luò)的學(xué)習(xí)速度，并且學(xué)習(xí)過程具有良好的收斂性及較強(qiáng)的魯棒性。

【關(guān)鍵字】神經(jīng)網(wǎng)絡(luò),算法,異或,學(xué)習(xí)率,反向傳播。

【Abstract】The XOR question can’t be implemented by the structure and study regulation of feeling machines, started with which, This production uses BP network to solve the XOR question, which obliterate the limitation of the feeling machine.But problems come into existence in concrete apply of BP network, for example: The restrained rate is slow, and there are existence of the strong coupling relation with other parameters. Otherwise the partial is extreme minute. So according to the mechinism of Feedforward Neural Network, this product put forward a method with variable learning rate factors for the improvement of BP algorithm. The inproved algorithm is applied to the learning of two or more dimensions XOR question.The simulations show the improved algorithm has good effects on speeding up learning process and bettering its learning convergence and robust performance.
【Key words】neural networks，algorithm，XOR，learning rate, back-propagation.
參考文獻(xiàn)
[1] 叢爽.面向matlab工具箱的神經(jīng)網(wǎng)絡(luò)理論與應(yīng)用.合肥：中國(guó)科學(xué)技術(shù)大學(xué)出版社.1998.11
[2] 王偉.人工神經(jīng)網(wǎng)絡(luò)原理——入門與應(yīng)用.北京：北京航空航天大學(xué)出版社.1995.10
[3] 武妍,張立明. 兩種將任務(wù)學(xué)習(xí)與模型學(xué)習(xí)相結(jié)合的神經(jīng)網(wǎng)絡(luò)習(xí)方法北京第屆學(xué)術(shù)年會(huì)論文. CNNC 12 , 2002-12
[4] 喬志駿,劉其真,易維列等. 一個(gè)基于模糊神經(jīng)網(wǎng)絡(luò)的數(shù)據(jù)逼近和泛化建模方法模式識(shí)別與人工智能 . 2001
[5] 宋宜斌，王培進(jìn). 多層前饋神經(jīng)網(wǎng)絡(luò)改進(jìn)算法及其應(yīng)用.計(jì)算機(jī)工程.2003.8