Probabilistic Modeling and Complex Systems Computation (CoSCo)

A complex system is a collection of simple interacting agents, elements or processes, whose collective behavior exhibits interesting large scale phenomena. Such systems can be found in various disciplines, including computer science, economics, mathematical biology and physics. The Complex Systems Computation (CoSCo) research group studies computational issues related to complex systems focusing on prediction and model selection issues. Current work of the CoSCo group is concentrated on theory and applications of Bayesian (belief) networks, and related probabilistic model families, such as finite mixture models. Other research areas addressed include case-based reasoning, artificial neural networks, and stochastic optimization methods, such as simulated annealing and genetic algorithms. The results achieved include extensive theoretical and empirical studies concerning

• efficient methods for learning probabilistic models from sample data;
• a novel, computationally efficient Bayesian criterion for selecting the most relevant model features (attributes);
• the accuracy of different marginal likelihood approximation techniques for Bayesian networks with hidden variables;
• similarities and differences between Bayesian and information-theoretic (MDL, MML) modeling approaches;
• a general Bayesian framework for case-based reasoning;
• techniques for mapping Bayesian networks to neural network architectures, thus allowing massively parallel implementations of Bayesian reasoning;
• computationally efficient stochastic optimization methods, including a novel variant of simulated annealing with an adaptive cooling schedule.

The theoretical results obtained have been empirically tested by using several real-world data sets. Some of the JAVA software used in these experiments can be downloaded from the group home page (http://www.cs.helsinki.fi/research/cosco/ ). Examples include

• BAYDA: Bayesian Predictive Discriminant Analysis with feature selection.
• D-SIDE: a Bayesian decision support system with JAVA interface.

In the empirical tests with various public domain classification datasets, D-SIDE consistently outperforms the results obtained by alternative approaches, such as decision trees and neural networks.

Basic research work by the group has been supported by grants from the Academy of Finland, University of Helsinki, and various foundations. More applied work has been performed with support from TEKES and the domestic and foreign industrial partners which include, e.g., Kone, Nokia, ABB, Enso and the AT&T corporations. Some of the resulting software has been adopted in the industry.

Current members of the CoSCo group are Doc. Henry Tirri, Dr. Petri Myllymäki, M.Sc. Tomi Silander, Petri Kontkanen, Jussi Lahtinen and Kimmo Valtonen. The group has also hosted several foreign academic visitors and graduate students, and participates in two EC-funded research networks (the NeuroCOLT working group on neural and computational learning theory, and the Network of Excellence in Neural Computing, NEURONET ).

Publications: [118, 120-128, 192, 193, 217, 223-233, 236, 237].

Home Page: http://www.cs.helsinki.fi/research/cosco/