You are here

Warning message

Attention! This event has already passed.

Learning causal models of multivariate systems and its value for the performance modeling of computer programs

Wednesday, 19 December, 2007 - 17:00
Campus: Brussels Humanities, Sciences & Engineering campus
Jan Lemeire
phd defence

Jan Lemeire's multidisciplinary work originated by the idea to introduce the causal learning algorithms into the world of performance analysis. Besides this practical and applied research it also comprises a theoretical and philosophical study of causal inference.

A performance analysis aims at understanding the execution of program on computer systems in terms of resource utilization. The causal learning algorithms allow the automatic construction from experimental data of models showing which and how variables influence the overall performance.

The feasibility to learn the causal mechanisms of a system from observations only is an ambituous, yet controversial subject.

The new concept of Kolmogorov Minimal Sufficient Statistic (KMSS) provides means to evaluate the validity of causal inference. The idea is that a model should capture all patterns or regularities of the observations.

A red thread throughout this work is the notion of qualitative properties. The proposal is to define them as the regularities of the KMSS, the properties of the data which allow compression. Qualitative properties provide another kind of knowledge than pure quantitative information. They are indispensable in the understanding of the system.