Structure learning for piecewise stationary varying DBN in model section

To learn the dynamic Bayesian network (DBN) structure under the limited sample data capacity and prior assumptions, the modeling approach for piecewise stationary and varying DBN in non-stationary stochastic process is studied. In the model section, the approximate representing by a first-order conditional independent DBN is utilized, which makes the model topology parse and leads to fast learning. An improved Markov chain Monte Carlo (MCMC) optimization algorithm for DBN structure learning is proposed, which avoids the pre-convergence in classical MCMC algorithm via increasing the Markov chain number adaptively. Comparison experimental results illustrate that the presented algorithm is more effective than classical MCMC or structural expectation maximization methods.