A Look-Ahead Simulation Algorithm for DBN Models of Biochemical Pathways
Résumé
Dynamic Bayesian Networks (DBNs) have been proposed [16] as an
efficient abstraction formalism of biochemical models. They have been shown to
approximate well the dynamics of biochemical models, while offering improved
efficiency for their analysis [17,18]. In this paper, we compare different representations
and simulation schemes on these DBNs, testing their efficiency and accuracy
as abstractions of biological pathways. When generating these DBNs, many
configurations are never explored by the underlying dynamics of the biological
systems. This can be used to obtain sparse representations to store and analyze
DBNs in a compact way. On the other hand, when simulating these DBNs, singular
configurations may be encountered, that is configurations from where no transition
probability is defined. This makes simulation more complex. We initially
evaluate two simple strategies for dealing with singularities: First, re-sampling
simulations visiting singular configurations; second filling up uniformly these
singular transition probabilities. We show that both these approaches are error
prone. Next, we propose a new algorithm which samples only those configurations
that avoid singularities by using a look-ahead strategy. Experiments show
that this approach is the most accurate while having a reasonable run time.
Domaines
Systèmes et contrôle [cs.SY]Origine | Fichiers produits par l'(les) auteur(s) |
---|
Loading...