一种通过节点序寻优进行贝叶斯网络结构学习的算法

Learning Bayesian Network Structure from Node Ordering Searching Optimal

针对K2算法过度依赖节点序,遗传算法节点序寻优效率差的问题,该文提出一种直接对节点序进行评分搜索的贝叶斯结构学习算法。该算法以K2算法为基础,首先通过计算支撑树权重矩阵,构建能够定量评价节点序的适应度函数。然后通过提出混合交叉策略和孤立节点处理机制,同时利用动态学习因子和倒置变异策略,提升遗传算法节点序寻优的性能。最后将得到的节点序作为K2算法的先验知识得到最优贝叶斯网络结构。仿真结果表明,该方法解决了K2算法依赖先验知识的问题,相比于其它优化算法,评分值平均增加了13.11%。

The performance of the K2 algorithm depends on node ordering heavily, and the genetic algorithm can not find the node ordering effectively. For these problems, a new Bayesian structure learning algorithm, named NOK2 （Node Ordering searching for K2 algorithm）, is proposed to solve the Bayesian structure learning problem by searching node ordering directly. According to the requirements of K2 algorithm based on prior knowledge and the weight matrix of spanning tree, the fitness function for quantitative evaluation of node ordering is established. The genetic algorithm is redesigned by a new method combines the dynamic learning constants, the hybrid crossover strategy, the inverted mutation strategy and the isolated node processing, so that the algorithm can find the node order of the highest fitness value, and this node sequence is taken as a prior knowledge of the K2 algorithm to obtain the optimal Bayesian network structure. Compared with other optimization algorithms, experimental results indicate that the NOK2 algorithm can significantly increase nearly 13.11% in the scoring metric values.