|
|
Project Title
Probabilistic
Network Library
Brief Description
Probabilistic Network Library (PNL) is the general tool
for working with graphical models. The library contains high-performance
implementation of algorithms for working with Bayesian networks and Markov
networks, such as belief propagation and Junction tree inference, maximum
likelihood and expectation maximization. The library is aimed at a wide spectrum
of graphical models applications including computer vision, pattern recognition,
data mining, and decision theory. The PNL core engine will be optimized and
parallelized to give maximum performance on Intel®
Architectures.
Research Area Description
The
probabilistic Network theory is one of the
most upcoming scientific areas over numerous
applications in bioinformatics, genetics, health protection, and
in multifarious areas of the computer science. Library
contains all the most well-known algorithms of teaching and
conclusion, so, undoubtedly, its development will be very good
for plentiful researchers, not only in the area of Probabilistic
Network
, and, above all, for specialist,
which occupied with various applications.
Goals
-
To
design, develop and implement scalable parallel versions of existing inference
and learning algorithms.
-
To
implement new types of probabilistic networks and distributions, inference and
learning algorithms for them.
-
To
optimize several algorithms from PNL.
-
To
provide connection or interface between PNL and some well-known tools of
probabilistic modeling.
Research Team
-
Gergel
V.P.
-
Belov
S.A.
-
Sysoyev
A.V.
-
Abrosimova
O.N.
-
Bader
A.A.
-
Vinogradov
R.V.
-
Gergel
A.V.
-
Labutina
A.A.
-
Senin
A.V.
-
Sidorov
S.V.
-
Tarasov
V.A.
-
Chernishova
E.N.
Main Results
-
Scalable parallel versions of inference algorithms
Junction Tree Inference, Loopy Belief Propagation, Gibbs Sampling and learning
algorithm Expectation Maximization Learning was developed and implemented.
-
Support of LIMID and diagnosis networks was added to
PNL including inference algorithm for them. SoftMax and Decision Tree
distributions were implemented, some inference algorithms were enlarged to
support networks with nodes that have these distributions.
-
Some matrix operations were optimized. Junction Tree
Inference algorithm was optimized too.
-
Connection
with GeNIe and interface with R were
implemented.
Current research
Publications
- Абросимова О.Н., Белов С.А., Сысоев А.В. Балансировка вычислительной
нагрузки для параллельных алгоритмов на вероятностных сетях. Материалы
третьего Международного научно-практического семинара “Высокопроизводительные
параллельные вычисления на кластерных системах”, 2003, 230-234
- Chernyshova E.N., Gergel A.V., Sysoyev A.V. Parallelization principles of
message passing algorithm for probabilistic networks, VI International
Congress on Mathematical Modeling/Book of abstracts, 2004, 38
- Belov S. Gergel V., Sysoev A., Scalable parallel inference algorithms in
probabilistic networks, Preproceedings of UK-Russia Workshop on Proactive
Computing, Nizhny Novgorod, February 2005 pp. 5-10
Conferences, seminars
- III Международный научно-практический семинар “Высокопроизводительные
параллельные вычисления на кластерных системах”, 2003, Н. Новгород.
- VI International Congress on Mathematical Modeling, 2004, N. Novgorod.
- UK-Rissia Workshop on Proactive Computing, Nizhny Novgorod, February 2005
pp. 5-10
Educational Materials
Developed software
|
| |
News
17.08.2005
17.08.2005
17.08.2005
17.08.2005
17.08.2005
23.05.2005
|