ACM Computing Surveys 28(4es), December 1996, http://www.acm.org/pubs/citations/journals/surveys/1996-28-4es/a4-doyle/. Copyright © 1996 by the Association for Computing Machinery, Inc. See the permissions statement below. This article derives from a position statement prepared for the Workshop on Strategic Directions in Computing Research.

Jon Doyle

---

Abstract: To survive and thrive, whether intellectually or as a force in industry, artificial intelligence (AI) must better identify its unique contributions. Failing this, it faces disrespect, decay, and dissolution. I believe one obtains a better identification both by dividing AI's subjects of study into the categories of rational psychology, psychological engineering, and articulating intelligence than by highlighting AI's characteristic methods of seeking computational-complexity explanations and simulating extremely complex semi-numerical models.

---

Major fields cannot survive without major problems or methods unique unto themselves. AI's recent and traditional difficulties lie in lacking such unique problems and methods, at least in the view of outsiders. The field has never agreed on its own definition, other than by stipulating that it traditionally encompasses several subjects, notably ``making intelligent machines'' and ``understanding intelligence''. These problems, however, do not possess the uniqueness requisite to a healthy discipline.

• The problem of understanding intelligence sits too close to the older fields of psychology and ethology (even theology), if taken to mean understanding intelligences we observe. This problem becomes unique only when taken to mean understanding possible intelligences other than those we observe.
• The problem of making intelligent machines sits too close to the more general activity of making more useful and capable machines, as reflected in the truism that successful AI systems generally no longer count as AI. The field has proven most successful at mechanizing isolated functions of intelligence, but differentiating this activity from that of building special machines generally appears hair-splitting at best and impossible at worst. The problem becomes unique only when taken to mean making machines that reason.

Even if we abandon these superficial but traditional ``mission statements'' and look more deeply, we still see that AI shares most of its major intellectual problems with other fields: understanding knowledge, reasoning, and rationality with logic, philosophy, psychology, economics, and sociology; understanding perception, motion, and manipulation with physiology, anatomy, and psychology; understanding language with linguistics and philosophy; understanding planning with economics and operations research. Surely no important aspect of what makes humans special has remained foreign to other fields, for---knowledge aside---people today have the same functions and abilities they have had for thousands of years.

Looking back at the history of the field, the methods of AI appear more distinctive than the problems addressed.

• The first new method employed by AI consisted of examining traditional problems by constructing and using symbolic or semi-numerical models of a complexity and character never seen before. Traditional modeling methods stretched to consider small sets of differential equations or linear equations of at most a couple thousand dimensions, but the nonlinear systems represented by computer programs of moderate size far outstrip these predecessors in complexity and potential difficulty of analysis. The name ``complex information processing'' early favored by Newell and Simon for AI suffers from understatement in capturing the essence of the advance represented by this method of the field.
• AI's second new method, inspired by the computational models, consisted of seeking explanations for psychological phenomena and justifications for putative psychological structures in terms of the notion of computational complexity, explaining observed limitations on the basis of the difficulty of computing something with a Turing machine (or equivalent device), and postulating one structure over another on the basis of computational advantages. The mechanistic assumptions presupposed by this method may have embroiled AI in philosophical and religious disputes, but the method has proved fruitful in specific psychological investigations quite independent of general suppositions about the nature of man.

For these reasons, AI must rethink itself and either identify problems unique to the field or give up and go home to the traditional fields. I entertain the possibility that the option of dissolution might prove right, but believe that AI should first try reviewing its purpose or purposes. I propose to abandon the expiring patent on the methods of AI and to find truly unique roles for AI along a cleavage induced by AI's broad conception of intelligence.

I see AI's most dramatic break with the past not in its computational methods but in its conception of intelligence divorced from embodiments in the people and creatures we find around us. Rather than limit attention to what already exists, AI contemplates what might, and views human and other extant intelligences as particular forms of a broader universal. No field before AI has studied the question of understanding intelligence construed this broadly, though philosophers and science-fiction writers speculate on special possibilities. The first step to finding uniqueness in AI lies in distilling out the problems related to intelligence broadly construed. Do this, and two natural fields emerge.

• The first field, rational psychology, addresses the problem of understanding the full range of possible psychologies---by which I mean possible organizations for minds---and classifying them according to their structures and properties. Though others might differ, I view the problems of understanding and classifying possible psychologies as an essentially mathematical problem. Rational psychology seeks to find the most appropriate concepts with which to characterize and describe psychologies, so as to understand the nature of and connections between psychological concepts. It does not presuppose or require any sort of rationality of the psychologies under investigation. I adopt the term ``rational psychology'', an old term used by Kant and James to mean philosophical psychology, in analogy with rational mechanics, the mathematical or conceptual investigation of mechanics. While the name has not spread much, the idea certainly has correspondents in some subsequent activities. I originally suggested this conception as the proper conception for what some call ``cognitive science'', but I later came to think that term more apt for the broader, not entirely mathematical subject. More recently, Glymour and others have employed the term ``android epistemology'' to cover all possible psychologies, but that term, in conflict with its advertised meaning, suggests a much narrower conception of knowledge alone in human-like beings. I thus prefer the broader, more descriptive, and more traditional term.
• Psychological engineering addresses the problem of constructing psychologies that exhibit specified properties, of finding economical designs for implementing or mechanizing agents with specified capacities or behaviors. This problem characterizes much work in AI, and represents the general engineering problem specialized to psychologies rather than to mechanisms, chemicals, genes, etc., hence the name psychological engineering.

AI might survive without rethinking itself, but only by virtue of increasingly hyperbolic advertising, or by the personal longevity of its adherents. But this sad fate seems unworthy of the gems obscured by the overburden of old tales about the aims and contributions of AI. To survive, AI must offer problems and methods other fields can respect, and it must share knowledge and technique with these fields when it also shares their problems. The latter transformation of AI, from intellectual isolate to conceptual trader, has been accelerating for some time, so its survival reduces to ensuring that it has something left at the end of this exchange. People may disagree about whether rational psychology and psychological engineering provide the right cleavage of AI's gems. But these fields do possess important, clearly understandable tasks all their own. I propose AI reset its work to expose these gems, leave its intellectual parents and cleave to this new joint identity.

Acknowledgments: I thank Joseph Schatz for valuable discussions and MIT for its support over the years. The position argued here restates one presented in papers of 1982--1994 listed below. The image of cleaving the field draws inspiration from George Miller's essay on ``dismembering cognition.''

Bibliography

1. Doyle, J., 1982. The foundations of psychology: a logico-computational inquiry into the concept of mind, CMU CSD, Report 82-149. Revised version published in Philosophy and AI: Essays at the Interface (R. Cummins and J. Pollock, eds.), Cambridge: MIT Press (1991), 39-77.
   • This paper develops the conception of studying all possible psychologies.
2. Doyle, J., 1983. What is rational psychology? toward a modern mental philosophy, AI Magazine, V. 4, No. 3, 50-53.
   • This paper reintroduces and explains the field of rational psychology.
3. Doyle, J., 1988. Big problems for artificial intelligence, AI Magazine, Vol. 9, No. 1, 19-22.
   • This editorial suggests essentially the same division of AI as proposed here.
4. Doyle, J., 1994. Reasoned assumptions and rational psychology, Fundamenta Informaticae, Vol. 20, No. 1 (Spring 1994).
   • This article provides an example of my own work in rational psychology.
5. Miller, George A., 1986. Dismembering cognition, in One Hundred Years of Psychological Research in America (Hulse, S. H. and B. F. Green, Jr., eds.). Baltimore: Johns Hopkins University Press, pp. 277-298.
   • Miller's essay focuses on how different theories of psychology divide the subject with different concepts, few of which cleanly ``divide the subject at its joints''.

Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, to republish, to post on servers, or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from Publications Dept, ACM Inc., fax +1 (212) 869-0481, or permissions@acm.org.