This article puts forward a solution named divide - assemble by deducing the size of bp neural network to overcome entering the local best point , the dividing process is that a big bp neural network is divided into several small bp neural networks , every small bp neural network can study alone , after all small bp neural networks finish their study , we can assemble all these small bp neural networks into the quondam big bp neural networks ; on the basis of divide - assemble solution , this article discusses the preprocessing of input species and how to deduce the size of bp neural network further to make it easy to overcome entering the local best point ; for the study of every small bp neural network , this article adopts a solution named gdr - ga algorithm , which includes two algorithms . gdr ? a algorithm makes the merits of the two algorithms makeup each other to increase searching speed . finally , this article discusses the processing of atm band - width distribution dynamically 本文從bp網(wǎng)的結(jié)構(gòu)出發(fā),以減小bp神經(jīng)網(wǎng)絡(luò)的規(guī)模為手段來克服陷入局部極小點,提出了bp神經(jīng)網(wǎng)絡(luò)的拆分組裝方法,即將一個大的bp網(wǎng)有機地拆分為幾個小的子bp網(wǎng),每個子網(wǎng)的權(quán)值單獨訓(xùn)練,訓(xùn)練好以后,再將每個子網(wǎng)的單元和權(quán)值有機地組裝成原先的bp網(wǎng),從理論和實驗上證明了該方法在解決局部極小值這一問題時是有效的;在拆分組裝方法基礎(chǔ)上,本文詳細闡述了輸入樣本的預(yù)處理過程,更進一步地減小了bp網(wǎng)絡(luò)的規(guī)模,使子網(wǎng)的學(xué)習(xí)更加容易了;對于子網(wǎng)的學(xué)習(xí),本文采用了最速梯度? ?遺傳混合算法(即gdr ? ? ga算法) ,使gdr算法和ga算法的優(yōu)點互為補充,提高了收斂速度;最后本文闡述了用以上方法進行atm帶寬動態(tài)分配的過程。