基于贝叶斯网络的权重自学习方法研究

Research on weights self-learning method based on Bayes Net

为了解决多层次、多属性、多目标复杂系统决策的权重确定问题,提出了一种首先对属性进行主观赋权,然后通过贝叶斯网络进行自学习的权重确定方法。贝叶斯网络的构造由决策目标之间的相互关系来确定,其条件概率表是属性相对于目标的初始权重和属性值所组成的二元结构体,其权重自学习梯度下降策略采用贪心爬山法。由于贝叶斯网络没有对输入层的结构要求,可以解决两层以上多目标决策权重的确定问题。用该方法来确定权重是通过网络对样本学习而获得的,学习过程不受人为因素的干扰,减少了复杂系统决策过程中的不确定性因素的影响,保证了多目标决策的可靠性和准确性。

A new method was put forward to resolve weights assignment problem of complex system with multi--layer, multi--attribute, multi--object, which firstly used subjective weight to assign attributes, and then adopted Bayes Net to learn the weights. Bayes Net structure was determined according to interrelationships among decision goals. Condition Probability Table was a two--unit framework composed of attribute value and initialized weight. Cupidity Mountain Climbing method was adopted to achieve grads descending strategy in weights self- learning method. Because the Bayes Net had no requirements on the structure of input layer, it could be used to solve weights assignment problem of multi--objective optimization with at least two layers. Weights assigned by this method were obtained through specimen learning by net. Learning process wasn＇t interfered by jamming, which reduced the influence of uncertain factors in the decision process of complex system so as to guarantee the reliability and veracity of multi--objective decision.