Estimation of state changes in system descriptions for dynamic Bayesian networks by using a genetic procedure and particle filters

Abstract

This paper deals with the estimation of state changes in system descriptions for dynamic Bayesian networks (DBNs) by using a genetic procedure and particle filters (PFs). We extend the DBN scheme to more general cases with unknown Directed Acyclic Graph (DAG) and state changes. First, we summarize the basic model of DBN where the DAG can be changed and the state transition occurs. In the genetic procedure to estimate DAG changes, we utilize the mutation operation (called Evolutionary Programming: EP) to the DAG to maintain consistency. By defining the possible DAG structure and state changes as particles, we formalize the optimization as the PF procedure. The weight of a particle representing the DAG and state transition is defined as the capability to approximate the probability distribution function obtained from a table of cases. We apply the estimation scheme of the paper to an artificially generated DBN, in which the state of the variables and the changed structure of the DAG are already known, to prove the applicability of the method, and discuss its applicability to debt rating.