Automating Path Analysis for Building Causal Models from Data.pdf

Automating Path Analysis for Building Causal Models from Data* Paul R. Cohen, Adam Carlson, Lisa Ballesteros, Robert St. Amant Computer Science Technical Report 93-38 Experimental Knowledge Systems Laboratory Department of Computer Science University of Massachusetts Amherst, MA 01003 Abstract Path analysis is a generalization of multiple linear regression that builds models with causal interpretations. It is an exploratory or discovery procedure for finding causal structure in correlational data. Recently, we have applied statistical methods such as path analysis to the problem of building models of AI programs, which are generally complex and poorly understood. For example, we built by hand a path-analytic causal model of the behavior of the Phoenix planner. Path analysis has a huge search space, however. If one measures N parameters of a system, then one can build O(2N2 ) causal models relating these parameters. For this reason, we have developed an algorithm that heuristically searches the space of causal models. This paper describes path analysis and the algorithm, and presents preliminary empirical results, including what we believe is the first example of a causal model of an AI system induced from performance data by another AI system. * This research was supported by DARPA-AFOSR contract F30602-91-C-0076. Automating Path Analysis. Cohen et al. 11. INTRODUCTION Because machine learning techniques search for structure in data, they often closely parallel statistical techniques for exploratory data analysis, such as clustering, factor analysis and regression. This paper describes a statistical discovery procedure for finding causal structure in corre- lational data, called path analysis [Asher, 83; Li, 75] and an algorithm that builds path-analytic models automati- cally, given data. This work has the same goals as re- search in function finding and other discovery techniques, that is, to find rules, laws, and mechanisms that underlie nonexperimen