An Improved Particle Swarm Optimization Algorithm for Bayesian Network Structure Learning via Local Information Constraint

Kun Liu,Peiran Li,Yani Cui,Jia Ren

doi:10.1109/access.2021.3065532

Abstract

At present, in the application of Bayesian network (BN) structure learning algorithm for structure learning, the network scale increases with the increase of number of nodes, resulting in a large scale of structure search space, which is difficult to calculate, and the existing learning algorithms are inefficient, making BN structure learning difficulty increase. To solve this problem, a BN structure optimization method based on local information is proposed. Firstly, it proposes to construct an initial network framework with local information and uses the Max-Min Parents and Children (MMPC) algorithm to construct an undirected network framework to reduce the search space. Then the particle swarm optimization (PSO) algorithm is used to strengthen the algorithm’s optimization ability by constructing a new position and velocity update rule and improve the efficiency of the algorithm. Experimental results show that under the same sample data set, the algorithm can obtain a more accurate BN structure while converging quickly, which verifies the correctness and effectiveness of the algorithm.

Highlights

Bayesian network (BN) is an important tool for uncertain knowledge expression, and it is a directed acyclic graph (DAG) of joint probability distribution of some nodes [1]
The DAG qualitatively represents the independent relationship between variables, and the conditional probability table (CPT) quantitatively represents the degree of dependence between variables
Aiming at the above problems, we propose a BN structure learning algorithm based on local information

Summary

INTRODUCTION

Bayesian network (BN) is an important tool for uncertain knowledge expression, and it is a directed acyclic graph (DAG) of joint probability distribution of some nodes [1]. Aiming at the above problems, we propose a BN structure learning algorithm based on local information This method first uses the local search-based Max-Min Parents and Children (MMPC) algorithm to construct an undirected network framework to optimize the initial population and reduce the search space of the PSO. Use the MMPC based on the restriction technology to construct the framework of the undirected network, thereby solving the problem of the large randomness of the initial particles in the later optimization process using the PSO algorithm. It carries the parents and children set information of the node, which improves the ability of the subsequent iterative optimization process of the PSO to jump out of the local extreme value. Initialize entire particle parameters, velocity matrix V and position matrix P

Update the position

Domain degree

TABLE IV EXPERIMENTAL RESULTS OF LIPSO PERFORMANCE FOR CONSTRUCTING

CONCLUSIONS