一类基于蚁群优化的贝叶斯置信网结构学习策略及性能分析

Performance of class of ant colony optimization based on Bayesian networks structural learning strategy

针对贝叶斯置信网的结构学习问题,提出一种遵循典型ACO算法框架(ACO-TSP)的贝叶斯网结构学习算法(ACO-BN),并拓展为包括EAS-BN、ACS-BN和MMAS-BN在内的一类算法。用这类算法在若干典型贝叶斯网络结构学习问题上分别与经典贝叶斯网学习算法(K2、B)、用于贝叶斯网学习的通用优化算法(simulated annealing、Tabu searching和genetic searching)以及L.M.de Campos等人提出的基于蚁群优化的贝叶斯网络结构学习算法Ant-K2SN和Ant-B进行了比较。实验结果表明,这类算法在总体性能上要优于经典贝叶斯网学习算法、通用优化算法以及Ant-K2SN和Ant-B算法,但是在时间性能上要略逊一筹。总的来说,这类算法是较为可行的一类贝叶斯置信网结构学习策略。

This paper presented an ACO Bayesian networks structural learning （BNSL） strategy （ACO-BN） follows the typical AGO framework （ACO-TSP） and extended to a class of ACO-BN algorithms including EAS-BN, ACS-BN and MMAS-BN. Experiments on some typical BNSL benchmark problems showed that ACO-BNs have much better capability than the classic BNSL algorithms （K2 and B）, general purpose optimization algorithms for BNSL （simulated annealing, Tabu searching and genetic searching） and the two ACO inspired algorithms called Ant-K2SN and Ant-B proposed by L. M. de Campos et. al in learning effective Bayesian networks structure. However, ACO-BNs have a little inferior time performance than the other alogrithms mentioned above. On the whole, the ACO-BNs can be regard as a kind of feasible BN structural learning strategy.