The current algorithms of learning the structure of dynamic Bayesian networks attempt to find single "best" model. However, this approach ignores the uncertainty in model selection and is prone to overfitting and local optimal problem. Markov chain Monte Carlo algorithm based on Bayesian model averaging can provide a way for accounting for this model uncertainty, but the convergence is too slow. Therefore, in this paper, a novel method, called DBN-EMC algorithm, is proposed which integrates techniques from evolutionary computation into the Markov chain Monte Carlo framework. In order to improve speed convergence, the algorithm introduces the mutation and crossover operation of the genetic algorithm to create new Markov chains to evolve the structure, and first gives a method to avoid the problem of the directed cyclic graph which is proposed by the mutation and crossover operation on the structure. The experimental results show the algorithm performs well. It not only can effectively learn the structure of dynamic Bayesian networks but also significantly improve the speed convergence of Markov chain Monte Carlo
[1]
Kevin P. Murphy,et al.
Learning the Structure of Dynamic Probabilistic Networks
,
1998,
UAI.
[2]
Thomas L. Dean,et al.
Probabilistic Temporal Reasoning
,
1988,
AAAI.
[3]
Wang Fei.
Research on Learning Dynamic Bayesian Networks by Genetic Algorithms
,
2003
.
[4]
J. York,et al.
Bayesian Graphical Models for Discrete Data
,
1995
.
[5]
Adrian E. Raftery,et al.
Bayesian model averaging: a tutorial (with comments by M. Clyde, David Draper and E. I. George, and a rejoinder by the authors
,
1999
.
[6]
Malcolm J. A. Strens,et al.
Evolutionary MCMC Sampling and Optimization in Discrete Spaces
,
2003,
ICML.
[7]
Nir Friedman,et al.
The Bayesian Structural EM Algorithm
,
1998,
UAI.
[8]
Kathryn B. Laskey,et al.
Population Markov Chain Monte Carlo
,
2004,
Machine Learning.