Abstract: Bayesian network structure learning algorithms based on scoring search need parameter tuning, which increases computational complexity. Improper parameters make the algorithm fall into local optimum. Without tuning, the Jaya algorithm is applied to Bayesian network structure learning. In the Jaya algorithm framework, combined with cross-variation of genetic algorithm, the individual update strategy is redesigned. Consequently, Jaya algorithm can be applied to the discrete structure learning optimization problem. The convergence of the proposed algorithm is discussed with the Markov chain theory. Experimental results show that the algorithm can be effectively applied to Bayesian network structure learning.

Key words: Bayesian network, structural learning, Jaya algorithm, Markov chain

摘要： 基于评分搜索的贝叶斯网络结构学习算法通常需要调参，导致计算量增大且不当的参数易使算法陷入局部最优。针对这一问题，将无需调参的Jaya算法应用于贝叶斯网络结构学习。在Jaya算法的框架下，结合遗传算法的交叉变异思想重新设计了个体更新策略，使Jaya算法能够应用于结构学习这一离散优化问题，并结合马尔科夫链的相关理论讨论了所提算法的敛散性。实验结果表明，该算法能有效应用于贝叶斯网络结构学习。

关键词: 贝叶斯网络, 结构学习, Jaya算法, 马尔科夫链