Browse Prior Art Database

LEARNING HIDDEN CAUSES FROM EMPIRICAL DATA

IP.com Disclosure Number: IPCOM000128345D
Original Publication Date: 1984-Dec-31
Included in the Prior Art Database: 2005-Sep-15
Document File: 12 page(s) / 36K

Publishing Venue

Software Patent Institute

Related People

Judea Pearl: AUTHOR [+3]

Abstract

This study is mativatCd by the obsCrvatioa that human beings, facing complex exhibit an almost obsessive urge to conceptually mold these phenomena into structures of cause-and-effect relationships. This tendency is, in fact, so oompulsivC that it sometimes comes at the expense of precision sad often requires the invention of hy_ pCth .ChC81, 11nObBCtiI8b1C entities such as "Cg0", "", and *supreme be-ings" to make thCOIlCS fit the mold of causal schema. When we try to 11adCrSIand the ac-tions of another person, for example, we invariably invoke abstract nations of mental states, social attitudes, beliefs, goals, plans and intentions. Medical knowledge, likewise, is organized into causal hiCrarchies ofinvading organisms, physical disorders, eomplim. boas, dinicai states, and, only finally, the visible symptoms. Ilk paper takes the position that homes obsession with causation is computation-lily motivated. Causal models are extractive because they provWe effective data-structurss far repr~atiag empirical knowledge, and their effeciivCaess is a remit of the high degree of dP mposition they induce. Man specifically, canna are viewed as names gtvCn to near variables which, once calculated, would permit us to treat the old variables as if they were independent. The deoompositional role of a causal variable is analogous to that of as orchestra coniducxor; it achieves coordinated behavior through central oaimmu~aication and thereby relieves the players from having to communicate directly with each other. Such ooordi-nation is characteristic of tree structures and draws its effectiveaess from the local na-ture of the data flow topology. In a management hierarchy. for example, w1xrC employ. ees can only communicate with each other through their immediate superiors, the pas-sage of information is swift, economical, conflict-free, and highly parallel. These com-putational attributes, we postulate, give rise to the satisfying sensation called "in-depth understanding", which people e:paience whey they discover causal models consistent with observations. Cast in probabilistic terms, central d~mposition is embodied by the relation of coafidonal lndepeadrnce, which we claim constitutes the most universal and distinctive characteristic feared by the notion of causality. (See also tea, 1952, and , 1970.) In medical diagnosis, for example, a group of co-occurring symptoms often bo-oomc independent of each other once we know the disease that caused them. When some of the symptoms diraxiy intluGace each other, the medical profession invents a name for that interaction (e.g., omaplication, clinical state, etc.) and treats it as a new auxiliary variable which again assumes the dooo~itionai role cbaractesistic of causal agents. Knowing the exact state of the auxiliary variable renders the interacting symp. tams independent of each other. Causes invoked to e:plaia human behavior, such as motives and intentions, also induce conditional indepen~. For e:nmple, once a murder suspect confesses to having wished the death of the victim, testimonies proving that he expressed such wishes in public or that he stood to gain fmm the victim's death are perceived to be irrelevant; they shed no further light on whirr he actually per-formed the murder.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 13% of the total text.

Page 1 of 12

THIS DOCUMENT IS AN APPROXIMATE REPRESENTATION OF THE ORIGINAL.

LEARNING HIDDEN CAUSES FROM EMPIRICAL DATA

Judea Pearl December 1984 Report No. CSD-840065

UCLA. ' November 1984 LEARNING ADEN CAUSES PROM Eh9iRICAL DATA

Judea Pearl

Cognitive Systems Laboratory Computer Sci Departmreot University of California Los Angeles, CA 90024 T: (213 825-3243 ~ '' Net address: jndea@UGLA LrOCUS.ARPA Topic: Learning

Keywords: Causal models, state variables, tree- dependence, Conftonalt adepeadeace

Ward Count: 3476 This work was supported is part by the National Science Foundation, Grant #DM 83-13875 _- ..

1. INTRODUCTION: CAUSALITY AND CONDITIONAL INDEPENDENCE

This study is mativatCd by the obsCrvatioa that human beings, facing complex exhibit an almost obsessive urge to conceptually mold these phenomena into structures of cause-and-effect relationships. This tendency is, in fact, so oompulsivC that it sometimes comes at the expense of precision sad often requires the invention of hy_ pCth .ChC81, 11nObBCtiI8b1C entities such as "Cg0", "", and *supreme be-ings" to make thCOIlCS fit the mold of causal schema. When we try to 11adCrSIand the ac-tions of another person, for example, we invariably invoke abstract nations of mental states, social attitudes, beliefs, goals, plans and intentions. Medical knowledge, likewise, is organized into causal hiCrarchies ofinvading organisms, physical disorders, eomplim. boas, dinicai states, and, only finally, the visible symptoms.

Ilk paper takes the position that homes obsession with causation is computation-lily motivated. Causal models are extractive because they provWe effective data-structurss far repr~atiag empirical knowledge, and their effeciivCaess is a remit of the high degree of dP mposition they induce. Man specifically, canna are viewed as names gtvCn to near variables which, once calculated, would permit us to treat the old variables as if they were independent.

The deoompositional role of a causal variable is analogous to that of as orchestra coniducxor; it achieves coordinated behavior through central oaimmu~aication and thereby relieves the players from having to communicate directly with each other. Such ooordi-nation is characteristic of tree structures and draws its effectiveaess from the local na-ture of the data flow topology. In a management hierarchy. for example, w1xrC employ. ees can only communicate with each other through their immediate superiors, the pas-sage of information is swift, economical, conflict-free, and highly parallel. These com-putational attributes, we postulate, give rise to the satisfying sensation called "in-depth understanding", which people e:paience whey they discover causal models consistent with observations.

Cast in probabilistic terms, central d~mposition is embodied by the relation of coafidonal lndepeadrnce, which we claim constitutes the most universal and distinctive characteristic feared by the notion of causality. (See also tea, 1952, and , 1970.) In...