The behavior of a distributed system or a network is subject to many irregularities and stochastic fluctuations. Our success in solving a variety of inference and optimization tasks defined over such systems depends heavily on our ability to adequately model, reason about and learn such a behavior.
Many existing stochastic models for complex systems and algorithms for their probabilistic analysis build upon the assumption of full independence; the condition that permits efficient probabilistic analysis but that is, at the same time, frequently violated in realistic world settings with intricate stochastic dependencies and interactions among components of the system. A simple analysis of real-world network systems (such as power grids, communication or transportation networks) reveals that situations with more failures occurring at the same time are more likely than in the model with failure independence in which an occurrence of a larger number of failures tends to be a very unlikely event. An example is August 2003 power blackout that affected large areas of Eastern USA and involved a cascade of interacting failures. Under the condition of failure independence the probability of such an event would be extremely small. The aim of our work is to develop probabilistic models of stochastic behaviors of complex distributed systems that overcome this difficulty, offer good approximations of true behavior of the system, and at the same time support efficient inferences and learning.
Research projects:
, PhD, Assistant
professor
T. Singliar and M. Hauskrecht. Noisy-or Component Analysis and its Application to Link Analysis. Journal of Machine Learning Research , 2006 (to appear).
T. Singliar and M. Hauskrecht. Variational Learning for the Noisy-OR Component Analysis. In the Proceedings of the SIAM International Data Mining conference, pp. 370--379 , 2005.
M. Hauskrecht, T. Singliar. Monte Carlo optimizations for resource allocation problems in stochastic network systems. In Proceedings of the Nineteenth International Conference on Uncertainty in Artificial Intelligence, pp. 305-312, 2003.
M. Hauskrecht, E. Upfal. A clustering approach to solving large stochastic matching problems. In Proceedings of the Seventeenth International Conference on Uncertainty in Artificial Intelligence, pp. 219-226, 2001.
T. Singliar and M. Hauskrecht. Modeling Highway Traffic Volumes. In Proceedings of Eighteen Europian Conference on Machine Learning (ECML), 2007, in press.
T. Singliar and M. Hauskrecht. Learning to detect traffic incidents from imperfectly labeled data. In Proceedings of Eighteen Europian Conference on Machine Learning (ECML), 2007, in press.
T. Singliar and M. Hauskrecht. Towards a learning traffic incident detection system. ICML 2006 Workshop on Machine Learning Algorithms for Surveillance and Event Detection, Pittsburgh, June 2006.
D. Mosse, L. Comfort, A. Labrinidis, A. Amer, J. Brustoloni, P. Chrysanthis, M. Hauskrecht, T. Znati, R. Melhem, K. Pruhs. Secure-CITI Project Highlights. Featured in the 7th Annual International Conference on Digital Government Research (dg.o 2006) San Diego, CA, May 2006.
The web page is updated by milos.