FROM BAYESIAN NETWORKS TO ACTIONS, CAUSES, AND COUNTERFACTUALS

Judea Pearl
Computer Science Department
UCLA

Abstract

Bayesian networks were introduced in the early 1980's to facilitate the representation of probability functions. Today, they are steadily giving way to a more flexible representation, called "modifiable structural model", which permits us to compute probabilities not merely for (conditional) events, but also for the effects of new actions (e.g., P(B=true)|do(A=true))), and for counterfactual sentences (e.g., "B would have been different if A were true"). The talk will describe the mathematical foundations of modifiable structural models, their axiomatic characterization, and their relations to the causal theories of Simon (1953), Good (1961), Suppes (1970) and Lewis (1973). Applications in AI, statistics, epidemiology, law, and economics will be demonstrated.

Reference material can be found in http://bayes.cs.ucla.edu/jp_home.html Technical Reports R-218, R-248, R-236T, R-240, R-250.


Eyal Amir
Last modified: Thu May 7 22:53:31 PDT 1998