Research on structure learning of dynamic Bayesian networks by particle swarm optimization

A new approach to learning structure of dynamic Bayesian networks (DBNs) is proposed in this paper. This approach is based on particle swarm optimization (PSO) algorithm. We start by giving a fitness function based on expectation to evaluate possible structure of DBNs by converting incomplete data to complete data using current best DBN of evolutionary process. Next, the definition and encoding of the basic mathematical elements of PSO are given and the basic operations of PSO are designed which provides guarantee of convergence. Next, samples for the incomplete training set and test set are generated from a known original dynamic Bayesian network with probabilistic logic sampling. Next, the structure of DBN is learned from incomplete training set using improved PSO algorithm steps. Finally, the simulation experimental results also demonstrate this new approach's efficiency and good performance in terms of predictive accuracy for test set.