基于超結(jié)構(gòu)的BN隨機搜索學習算法
發(fā)布時間：2019-05-07 01:55
【摘要】：近年來,貝葉斯網(wǎng)絡(Bayesian network,BN)在不確定性知識表示與概率推理方面發(fā)揮著越來越重要的作用.其中,BN結(jié)構(gòu)學習是BN推理中的重要問題.然而,在當前BN結(jié)構(gòu)的2階段混合學習算法中,大多存在一些問題:第1階段無向超結(jié)構(gòu)學習中存在容易丟失弱關系的邊的問題;第2階段的爬山搜索算法存在易陷入局部最優(yōu)的問題.針對這2個問題,首先采用Opt01ss算法學習超結(jié)構(gòu),盡可能地避免出現(xiàn)丟邊現(xiàn)象;然后給出基于超結(jié)構(gòu)的搜索算子,分析初始網(wǎng)絡的隨機選擇規(guī)則和對初始網(wǎng)絡隨機優(yōu)化策略,重點提出基于超結(jié)構(gòu)的隨機搜索的SSRandom結(jié)構(gòu)學習算法,該算法一定程度上可以很好地跳出局部最優(yōu)極值;最后在標準Survey,Asia,Sachs網(wǎng)絡上,通過靈敏性、特效性、歐幾里德距離和整體準確率4個評價指標,并與已有3種混合學習算法的實驗對比分析,驗證了該學習算法的良好性能.
[Abstract]:In recent years, Bayesian networks (Bayesian network,BN) play an increasingly important role in uncertain knowledge representation and probabilistic reasoning. Among them, BN structure learning is an important problem in BN reasoning. However, in the current two-stage hybrid learning algorithms of BN structure, there are some problems: the first stage of undirected superstructure learning has the problem of easily losing the weak relations; The second stage of mountain climbing search algorithm is prone to fall into the local optimal problem. In order to solve these two problems, firstly, Opt01ss algorithm is used to study the superstructure, so as to avoid the phenomenon of losing edges as far as possible. Then the search operator based on superstructure is given, the random selection rule of the initial network and the stochastic optimization strategy of the initial network are analyzed, and the learning algorithm of SSRandom structure based on the random search of superstructure is put forward emphatically. To some extent, the algorithm can jump out of the local optimal extremum. Finally, on the standard Survey,Asia,Sachs network, through the sensitivity, specificity, Euclidean distance and the overall accuracy of four evaluation indicators, and compared with the existing three hybrid learning algorithm of the experimental analysis, to verify the good performance of the learning algorithm.
【作者單位】： 山西大學計算智能與中文信息處理教育部重點實驗室;山西財經(jīng)大學信息管理學院;
【基金】：國家自然科學基金重點項目(61432011) 軍民共用重大研究計劃聯(lián)合基金重點項目(U1435212) 國家自然科學基金優(yōu)秀青年科學基金項目(61322211);國家自然科學基金項目(61672332) 中國博士后科學基金項目(2016M591409) 山西省自然科學基金項目(2013011016-4,2014011022-2)~~
【分類號】：TP18
本文編號：2470674
[Abstract]:In recent years, Bayesian networks (Bayesian network,BN) play an increasingly important role in uncertain knowledge representation and probabilistic reasoning. Among them, BN structure learning is an important problem in BN reasoning. However, in the current two-stage hybrid learning algorithms of BN structure, there are some problems: the first stage of undirected superstructure learning has the problem of easily losing the weak relations; The second stage of mountain climbing search algorithm is prone to fall into the local optimal problem. In order to solve these two problems, firstly, Opt01ss algorithm is used to study the superstructure, so as to avoid the phenomenon of losing edges as far as possible. Then the search operator based on superstructure is given, the random selection rule of the initial network and the stochastic optimization strategy of the initial network are analyzed, and the learning algorithm of SSRandom structure based on the random search of superstructure is put forward emphatically. To some extent, the algorithm can jump out of the local optimal extremum. Finally, on the standard Survey,Asia,Sachs network, through the sensitivity, specificity, Euclidean distance and the overall accuracy of four evaluation indicators, and compared with the existing three hybrid learning algorithm of the experimental analysis, to verify the good performance of the learning algorithm.
【作者單位】： 山西大學計算智能與中文信息處理教育部重點實驗室;山西財經(jīng)大學信息管理學院;
【基金】：國家自然科學基金重點項目(61432011) 軍民共用重大研究計劃聯(lián)合基金重點項目(U1435212) 國家自然科學基金優(yōu)秀青年科學基金項目(61322211);國家自然科學基金項目(61672332) 中國博士后科學基金項目(2016M591409) 山西省自然科學基金項目(2013011016-4,2014011022-2)~~
【分類號】：TP18
【相似文獻】
相關期刊論文 前2條
1 王永皎;米浩坤;梁磊;;應變和溫度變化對超結(jié)構(gòu)光纖光柵光譜的影響[J];光通信研究;2013年03期
2 李中科,楊曉輝,吳樂南;一種無需隨機搜索策略的自動平面檢測方法(英文)[J];Journal of Southeast University(English Edition);2003年03期
相關碩士學位論文 前1條
1 武佳杰;基于超結(jié)構(gòu)與隨機搜索的BN結(jié)構(gòu)學習算法[D];山西財經(jīng)大學;2017年
，本文編號：2470674
本文鏈接：http://sikaile.net/kejilunwen/zidonghuakongzhilunwen/2470674.html
最近更新
教材專著
