Dynamic Bayesian network structure learning based on an improved bacterial foraging optimization algorithm.

Area of Science:

• Artificial Intelligence
• Data Science
• Probabilistic Graphical Models

Background:

• Dynamic Bayesian Networks (DBNs) are widely used in engineering.
• Swarm intelligence algorithms offer robust optimization for DBNs.
• Existing methods struggle with DBN structure learning efficiency and accuracy.

Purpose of the Study:

• To propose an improved bacterial foraging optimization algorithm (IBFO-A) for enhanced global and local search capabilities.
• To develop a novel DBN structure learning method (IBFO-D) leveraging IBFO-A for complex, time-based data.
• To improve the efficiency and accuracy of DBN structure learning.

Main Methods:

• Developed IBFO-A with a four-layer framework: chaotic mapping initialization, Osprey-inspired exploration, genetic-based propagation, and elimination-dispersal.
• Introduced IBFO-D for DBN structure learning, integrating dynamic K2 scoring, V-structure rules, and trend activity.
• Applied "survival of the fittest" and "elimination-dispersal" strategies for optimal network structure selection.

Main Results:

• IBFO-A demonstrated good convergence, stability, and accuracy on benchmark functions and diverse data types.
• IBFO-D effectively learned DBN structures, outperforming existing methods in accuracy and efficiency.
• Experimental validation on benchmark test functions and 2T-BN networks confirmed the algorithms' performance.

Conclusions:

• IBFO-A provides a robust optimization framework for swarm intelligence.
• IBFO-D offers a practical and effective solution for DBN structure learning from data.
• The proposed methods have significant value for engineering applications involving complex temporal data analysis.