Learning Dynamic Bayesian Networks structure based on a new hybrid K2-Bat learning algorithm

Abstract

ABSTRACT The temporal dimension makes it difficult and complex to learn the Dynamic Bayesian Networks structure for huge search space. We propose a new hybrid K2-Bat algorithm to learn the structure of Dynamic Bayesian Networks. This work contains two optimal strategies: an ordering-based algorithm INOK2 to learn initial network structure and an adaptive binary bat algorithm to learn transition network structure. Based on the requirement of K2 algorithm for prior knowledge, a fitness function is built to quantitatively score node order in INOK2. The initial population is generated by the node block sequence constructed by directional support tree. A dynamic learning factor, inverted mutation s`node sequence to improve the global searching ability and the convergence speed. Then, the optimal initial network structure can be obtained. In addition, an improved binary bat algorithm is proposed to improve the development behavior of bat algorithm by using dynamic selection strategy in transitional network learning. Finally, experiments on four well-known benchmark problems are performed. The results show that the proposed algorithm can successfully learn the structure of Dynamic Bayesian Networks without prior knowledge, and balance solutions quality and computational effort.