Understanding Causal Mechanisms by the Controlled Lesion Method

Gregory D. Weber

My research suggests two new principles by which the causal structure of simple deterministic mechanisms may be inferred, based on controlled experiments, with very little background knowledge. The nmndom operador principle states that operators, or decision variables controlled by the agent, are not caused by, and do not share common causes with, state variables of the system being experimented with. This principle establishes an initial causal direction, pointing away from the operators tow.~rds the state variables. The lesion principle shows how, by exerting appropriate experimental control such as "cutting out" certain parts of the mechanism, the experimenter may infer a causal order (betweennem relation) among the variables. Together with Occam’s razor, these principlea enable an intelligent system to infer a directed, ordered cause-effect chain. They form the potential basis of a systematic method of inferring canssi relations from ezperiment. This method, called the Controlled Lesion Method, supplies the very background knowledge which was presupposed by the familiar notion described above.


This page is copyrighted by AAAI. All rights reserved. Your use of this site constitutes acceptance of all of AAAI's terms and conditions and privacy policy.