Extending Evolutionary Programming Methods to the Learning of Dynamic Bayesian Networks

Recent work has shown that for finding static Bayesian network structures, an Evolutionary Programming (EP) approach that exploits the description length of single links is better suited than a standard Genetic Algorithm (GA). We extend this work to find good dynamic Bayesian network structures that can have large time lags. We do this through the use of a new representation of dynamic Bayesian networks for EPs and a new operator, swap, designed specifically with a dynamic Bayesian network in mind. In this paper Lam’s knowledge guided operator for static networks is compared with the new swap operator when both are used in conjunction with the new representation. Experiments are carried out on synthetic datasets and a real world oil refinery process time series. The results indicate that the new operator is better suited to finding good structures in a shorter amount of time.