Interactive construction of decision models based on causal mechanisms

Authors:

Marek J. Druzdzel and Tsai-Ching Lu
Decision Systems Laboratory
School of Information Sciences
and Intelligent Systems Program
University of Pittsburgh
e-mail: marek@sis.pitt.edu, ching@isp.pitt.edu

Tze Yun Leong
Medical Computing Laboratory
Department of Information Systems and Computer Science
National University of Singapore
Singapore 119260
e-mail: leongty@iscs.nus.edu.sg

Abstract:

Quality of decisions based on the decision-theoretic approach depends on the quality of the underlying models. Construction of these models is outside of the realm of both probability theory and decision theory and is usually very laborious. Aiding model building in computer systems can significantly reduce the model construction time while increasing model quality and can contribute to a wider applicability of decision theory in decision support systems.

We propose an approach to computer-aided model construction that builds on the concept of causal mechanisms. Causal mechanisms are local interactions among domain variables that are usually fairly well understood and model independent, hence can be reused in different models. Their algebraic descriptions are known as structural equations. A model composed of causal mechanisms is causal and intuitive for human users. It also supports predictions of the effect of external interventions (decisions). We discuss issues related to storage and maintenance of causal mechanisms and interactive model building, including treatment of reversible causal mechanisms.