|
Abstract : |
Monitoring a complex dynamical system requires maintaining beliefs about its state. Even when processes are represented compactly as dynamic Bayesian networks, the computational complexity of exact monitoring is exponential in the number of state variables. Therefore tradeos must be made between the cost of computation and the precision of the inference via approximate monitoring. This paper presents a new family of approximate monitoring algorithms that combines the best qualities of the particle ltering and Boyen-Koller methods. Our algorithms maintain an approximate representation the belief state in the form of sets of factored particles, that correspond to samples of clusters of state variables. Empirical results show that the factored particles method outperforms ordinary particle ltering., |
