| Dynamic Bayesian Networks (DBNs) can serve as succinct models of large biochemical networks [19]. To analyze these models, one must compute the probability distribution over system states at a given time point. Doing this exactly is infeasible for large models and hence approximate methods are needed. The Factored Frontier algorithm (FF) is a simple and efficient approximate algorithm [25] that has been designed to meet this need. However the errors it incurs can be quite large. The earlier Boyen-Koller (BK) algorithm [3] can also incur significant errors. To address this, we present here a novel approximation algorithm called the Hybrid Factored Frontier (HFF) algorithm. HFF may be viewed as a parametrized version of FF. At each time slice, in addition to maintaining probability distributions over local states -as FF does- we also maintain explicitly the probabilities of a small number of global states called spikes. When the number of spikes is 0, we get FF and with all global states as spikes, we get the - computationally infeasible- exact inference algorithm. We show that by increasing the number of spikes one can reduce errors while the additional computational effort required is only quadratic in the number of spikes. We have validated the performance of our algorithm on large DBN models of biopathways. Each pathway has more than 30 species and the corresponding DBN has more than 3000 nodes. Comparisons with the performances of FF and BK show that HFF can be a useful and powerful approximation algorithm for analyzing DBN models of biopathways. |
Récupérer au format
