FREE ELECTRONIC LIBRARY - Abstracts, online materials

Pages:     | 1 |   ...   | 5 | 6 || 8 |

«Abstract Combinatorial explosion of inferences has always been a central problem in arti cial intelligence. Although the inferences that can be drawn ...»

-- [ Page 7 ] --

The conceptual dependency representation proposed by Schank and Abelson 1977] included D-KNOW, a goal to \change knowledge state," or to learn something. Examples of D-KNOW goals were to nd out the location of food (in order to go to it and then eat it) or to nd out the price of an item (in order to buy it). The generation of D-KNOW goals was always tied very speci cally to a physical supergoal (e.g., satisfy hunger), and were not mentioned in the author's later theories of learning (e.g., Schank, 1982]).

Other theories, particularly from the animal learning psychology literature, have proposed di use motivations to learn: a \will to perceive" (Thorpe), a \motivation for learning" (Thacker), and a \search by an information hungry organism" (Pribram, all reported in Livesey 1986], p. 20{21). As discussed above, an important feature of knowledge goals are their speci city. Merely desiring knowledge generally is not su cient for use in focussing attention, or in other important decisions during learning.

Social psychologists have used various \goal orientations" as explanatory phenomena in theories of attention, recall and judgement. These goal orientations are close in spirit to our knowledge goals. However, social psychologists' goal orientations are generally speci ed at a very abstract level (e.g., \form an impression," or \make predictions"). As Zukier's 1986] review notes, \In general, however, little systematic research is available on goal orientation in inference, and no comprehensive taxonomies of `middle-level' or concrete goals have emerged from these studies." Our work has described a much more detailed taxonomy of knowledge goal types, and proposed methods for their generation and use in processing.

Lenat's 1976] AM program had a method for focusing its attention that is related to our knowledge goals. AM maintained a queue of concepts to modify, and had a set of possible modi cations that could be applied. It chose which concept and which modi cation based on a heuristic evaluation of the interestingness of the concept. Each concept was tagged with an interestingness number, which was used to order the queue of concepts to change. AM's search of concept space was undirected; it was not trying to learn anything in particular, and therefore cannot be said to have goals for speci c knowledge. On the other hand, AM's interestingness heuristics contained an implicit characterization of what knowledge was desirable, for example, concepts with a few instances (not a single instance and not many instances). All of these characterizations were, however, syntactic; that is, they described the structure of interesting concepts, not their content. Nevertheless, this approach is more compatible with the use of knowledge goals than many systems, since, at least locally, the program made selections among various possible actions based on characterizations of the knowledge that would result from those actions.

Goal-directed planning has been investigated in the context of other reasoning tasks. For example, Tong's work on goal-directed planning in knowledge-based design addresses issues in task prioritization and inference control in the context of the design task Tong, 1987]. Tong's focusses on the development of knowledge-based models for design, and presents a framework for organizing, evaluating and developing such models from the perspectives of the knowledge embodied in the process model, the functionality of the design process, and the implementation of the design process as an actual program. In contrast, we are interested in developing goal-directed process models, as well as content theories, for learning across a variety of di erent reasoning tasks. Thus our work deals with cognitive domains rather than physical domains. Inference control in Tong's model is performed through the propagation of design constraints, and attempts to maintain consistency among speci cation details that must be true of a solution description. Our work focusses, not so much on the truth or correctness of inferences, as on the utility of these inferences for reasoning and learning.

Our process of planning to learn does not involve detailed reasoning about subproblem interactions (e.g., Sussman, 1975; Sacerdoti, 1975]) and constraint propagation (e.g., Ste k, 1981]) that is the focus of much work in conventional planning and problem solving. Our work is closer to research in opportunistic planning (e.g., Hayes-Roth and Hayes-Roth, 1979; Hammond, 1988]), and deals with the opportunistic pursuit of cognitive goals (e.g., Birnbaum and Collins, 1984; Dehn, 1989]), and the use of such goals to focus inference and learning. In order to incorporate a complete model of knowledge planning using learning actions, we are currently investigating issues such as prioritization of knowledge goals, and contention for resources such as time and storage space. Although knowledge goals do not directly interfere with each other (e.g., learning one thing does not disable the preconditions for learning another), we expect there will be signi cant interactions between goals due to resource constraints. The extent to which least commitment approaches to goal-directed planning in physical domains (as discussed, e.g., by Tong 1987]) will generalize to planning with knowledge goals in cognitive domains remains an open issue.

Also related to our claims is the work of Horvitz, et al 1989]. They present a calculus for deciding when to do more inference (versus when to act) in medical decision making. Although based on highly idealized functions for estimating the expected value of additional inference (in their model, inference includes data gathering), it provides an attempt to model content-based decisions about when it is worthwhile to acquire knowledge. Although their model does not specify what is worth learning, it may be useful in deciding whether it is worth learning at all, potentially reducing the size of the potential hypothesis space to zero.

Minton 1988] also proposes a model of judging whether it is worth learning, although his model involves computing the e ect of learning on future performance after the new concept is formed, and is hence not useful for focussing attention.

8 Conclusion: Automating curiosity We view learning as an incremental process of belief formation, involving a wide variety of interrelated learning processes. Most learning systems have incomplete knowledge of their domains; these gaps give rise to di culties during processing, and we propose that the di culties should give rise to explicit motivations to learn. Such a system learns in an incremental manner, by noticing interesting aspects of its experiences, generating knowledge goals based on those observations, and devoting some of its resources to achieving those goals. The process of satisfying those goals generally involves the selection of both an appropriate learning method and the focussing of attention on potentially relevant information sources.

Knowledge goals specify both desired knowledge and what to do with that knowledge once it is found.

The use a new piece of knowledge is put to (or, similarly, where that knowledge is stored in memory) depends on the motivation for acquiring that knowledge in the rst place. In AQUA, a new piece of knowledge could result in a new explanation in memory; it could be used to ll in a gap in an existing explanation; it could be used to elaborate an existing explanation if that explanation was not detailed enough to deal with the new situation; or it could be used to reorganize or re-index knowledge in memory to allow the reasoner to use what it already knows in novel situations to which that piece of knowledge had not been applied before.

In IVY, a new piece of knowledge can change provide speci c information to replace a general weak method for recognition, speci cation or distinction; it can generate a new subclass of a disease type in memory; it can be used to change the perceived value of gathering other pieces of information during diagnosis; or it can also cause the augmentation or reorganization of existing knowledge of indicative features of a disease or other entity. Each type of learning leaves the system a little closer to a complete understanding of its domain. Each type of learning can also result in a new set of knowledge goals. The satisfaction of one goal can lead to the identi cation of many other pieces of information that have become useful as a result.

An important class of knowledge goals for AQUA are those that are intended to test hypotheses that the program has generated. Hypotheses often have questions attached to them, representing what is still not understood or veri ed about those hypotheses. As the program reads new stories, it is reminded of past cases, and of old explanations that it has tried. In attempting to apply these explanations to the new situation, it also remembers unanswered questions that it had thought of previously. The system's understanding of its cases gradually gets re ned as these questions get answered. Details of this process may be found in Ram 1989; 1990b; 1992].

Much human learning seems subjectively to be a process of this type. Adults learn by modifying what they already know, using little pieces of new information as they come along. They have topics that they are interested in, and that they expend energy to pursue. People who are always asking new questions, and always on the lookout for new knowledge, are termed curious.

What would it take for a computer program to be curious? Any system that asks questions or gathers data about the world can be said to be curious in a very basic way: namely, it acts to acquire information.

This kind of behavior is not a very interesting model of curiosity. Even people who are considered gluttons for knowledge do not infer everything that can be inferred; they focus on particular aspects of their environment.

There is a well known psychiatric case of a person who had immense recall, but did not distinguish between relevant and irrelevant material Luria, 1968]; his pathology is not considered a kind of wild curiosity. Humanlike curiosity seems to us to require motivated pursuit of knowledge, or active learning, and not just the simple absorption of data and their consequences. Curiosity involves speci c (although often abstract) desires for knowledge, not merely a di use drive of some sort. When to attribute goals to computer programs is a di cult philosophical question (see, e.g., Dennett, 1987]), but we believe that programs that make decisions about what to learn and how to learn it have taken an important step toward genuine automated curiosity.

Our intent was not to have AQUA or IVY learn the \correct" understanding of terrorism or lung cancer, but rather to be able to wonder about unusual things they encounter and be motivated by those encounters to seek out new information. As they learn more about their domains, they ask better and more sophisticated questions. We suggest that both programs can be seen as simple models of human-like curiosity, and propose that the more practical or functional reasons for goal-directed learning processes, as discussed in this paper, be viewed as an explanation for the utility of this type of behavior.

The theory of knowledge goals presented in this paper brings together both cognitively motivated processes and functionally justi ed resource constraints, and provides a basis for designing practical reasoning systems that can represent and reason about their own goals.

Acknowledgements Ashwin Ram's research was supported in part by the National Science Foundation under grant IRI-9009710.

Part of the research described was conducted while Dr. Ram was at Yale University, and supported by the Defense Advanced Research Projects Agency and the O ce of Naval Research under contract N00014-85K-0108, and by the Air Force O ce of Scienti c Research under contracts F49620-88-C-0058 and AFOSRReferences Birnbaum and Collins, 1984] L. Birnbaum and G. Collins. Opportunistic Planning and Freudian Slips. In Proceedings of the Sixth Annual Conference of the Cognitive Science Society, pages 124{127, Boulder, CO,

1984. Institute of Cognitive Science and University of Colorado, Boulder.

Cox and Ram, 1991] M. Cox and A. Ram. Using Introspective Reasoning to Select Learning Strategies. In R. S. Michalski and G. Tecuci, editors, Proceedings of the First International Workshop on Multi-Strategy Learning, pages 217{230, Harpers Ferry, WV, November 1991. Center for Arti cial Intelligence, George Mason University, Fairfax, VA.

Dehn, 1989] N. Dehn. Computer Story Writing: The Role of Reconstructive and Dynamic Memory. Ph.D.

thesis, Yale University, Department of Computer Science, New Haven, CT, 1989. Research Report #792.

DeJong, 1979] G. F. DeJong. Skimming Stories in Real Time: An Experiment in Integrated Understanding.

Ph.D. thesis, Yale University, Department of Computer Science, New Haven, CT, May 1979. Research Report #158.

Dennett, 1987] D. Dennett. The Intentional Stance. Bradford Books/MIT Press, Boston, MA, 1987.

Dietterich, 1989] T. G. Dietterich. Limitations on Inductive Learning. In Proceedings of Sixth International Workshop on Machine Learning, pages 125{128, Ithaca, NY, June 1989. Morgan Kaufman.

Doyle, 1979] J. Doyle. A Truth Maintenance System. Arti cial Intelligence, 12:231{272, 1979.

Dyer, 1982] M. G. Dyer. In-Depth Understanding: A Computer Model of Integrated Processing for Narrative Comprehension. Ph.D. thesis, Yale University, Department of Computer Science, New Haven, CT, May

1982. Research Report #116.

Hammond, 1988] K. J. Hammond. Opportunistic Memory: Storing and Recalling Suspended Goals. In J. L. Kolodner, editor, Proceedings of a Workshop on Case-Based Reasoning, pages 154{168, Clearwater Beach, FL, May 1988. Morgan Kaufmann, Inc., San Mateo, CA.

Hayes-Roth and Hayes-Roth, 1979] B. Hayes-Roth and F. Hayes-Roth. A Cognitive Model of Planning.

Cognitive Science, 2:275{310, 1979.

Hayes-Roth and Lesser, 1976] F. Hayes-Roth and V. Lesser. Focus of attention in a distributed logic speech understanding system. In Proceedings of the IEEE International Conference on Accoustics, Speech and Signal Processing, pages 416{420, Philadephia, PA, April 1976. IEEE, New York, NY.

Hidi and Baird, 1986] S. Hidi and W. Baird. Interestingness | A Neglected Variable in Discourse Processing. Cognitive Science, 10:179{194, 1986.

Pages:     | 1 |   ...   | 5 | 6 || 8 |

Similar works:

«Minor illness, minor injuries and when to see your doctor. This booklet is provided by the Christmas Maltings Surgery to help you deal with minor illnesses yourself. It aims to explain some minor illnesses and injuries, suggest some simple remedies, and offer guidance about when you might need medical help. What do you mean, a minor illness? It doesn’t feel minor! A minor illness usually means one that will get better by itself, without causing any lasting harm. It doesn’t mean it feels...»

«Community Benefit Report Hoag Memorial Hospital Presbyterian OSHPD Facility ID #106301205 Gwyn P. Parry, M.D. Director, Community Health Telephone: (949) 764-5724 Email: Gwyn.Parry@hoag.org Minzah Z. Malik, MPH, MBA Manager, Community Benefit Program Telephone: (949) 764-6597 Email: Minzah.Malik@hoag.org Hoag Memorial Hospital Presbyterian www.hoag.org Board of Directors Hoag Memorial Hospital Presbyterian Fiscal Year 2011 Officers Stephen Jones, Chairman Robert W. Evans, Vice Chairman John L....»

«ACHS Perfumery Manual PLEASE NOTE The information in this manual is not intended to take the place of diagnosis and treatment by a qualified licensed health care provider. Any recommendations are for educational purposes only and are believed to be effective. However, since use of any material by others is beyond the control of Australasian College of Health Sciences, no expressed or implied guarantee as to the effectiveness of this information can be given nor liability taken. LAST REVISION...»

«International Cancer Screening Network 2015 June 2 – 4, 2015 • Rotterdam, The Netherlands Biosketches Co-chairs Harry de Koning, Ph.D. Dr. de Koning is a Professor of Public Health and Screening Evaluation at Erasmus University in Rotterdam, The Netherlands. Dr. de Koning's major scientific contributions are in the areas of (1) designing, running, and evaluating large-scale, multidisciplinary, population-based, randomized controlled screening trials to establish the efficacy of screening;...»

«Wiki Guide PDF Battlefield 3 Multiplayer Multiplayer Tips Basics Using Kits Playing Defensive Support Effectively Using Weapons Advanced Tactics Countering mortar fire as a team Squads Vehicle Tips Body part damage multipliers Health Experience Points Leveling Up Guide Basic Bonuses Squad Bonuses Kit Bonuses Vehicle Bonuses Objective Bonuses Ribbons Medals Service Stars Unlock Trees Ranks Kits (Unlocks) Vehicles (Unlocks) Multiplayer Modes Conquest Rush Squad Rush Team Deathmatch Squad...»

«Package leaflet: Information for the user Targin® 5 mg/2.5 mg, 10 mg/5 mg, 20 mg/10 mg and 40 mg/20 mg prolonged-release tablets Oxycodone hydrochloride and Naloxone hydrochloride Read all of this leaflet carefully before you start taking this medicine because it contains important information for you.  Keep this leaflet. You may need to read it again.  If you have any further questions, ask your doctor or pharmacist.  This medicine has been prescribed for you only. Do not pass it on...»


«Using a Phase I Environmental Site Assessment to Document Compliance with HUD Environmental Standards at 24 CFR 58.5(i)(2) or 50.3(i) Introduction HUD policy requires that “all properties that are being proposed for use in HUD programs be free of hazardous materials, contamination, toxic chemicals and gases, and radioactive substances, where a hazard could affect the health and safety of occupants or conflict with the intended utilization of the property.”1 Site contamination hazards must...»

«ABC LAW FIRM ATTORNEYS AND COUNSELLORS January 23, 2015 FILE #: 1167.0002 Confidential and Privileged Communication 1 Attorney Work Product/Attorney Client and Quasi Attorney Client Communication to Insurer Donna Livingston, Asst. VP Health Care Claims OneBeacon Professional Insurance Via Electronic Copy Style of Case: Venue: Claim Number (One Beacon): 0ab042998 Dear Donna, Pursuant to our previous discussions, this Case Evaluation Report is provided. Overview of Claim This is a wrongful death...»

«community. Welcome to the Arizona Children’s Rehabilitative Services (CRS) Member Handbook © 2016 United Healthcare Services, Inc. All rights reserved. CSAZ16MC3840379_000 Important Information Member Services: Available 8:00 a.m. – 5:00 p.m. Monday – Friday excluding state holidays. Toll-Free........................................................ 1-800-348-4058 TDD (for the hearing impaired).......................»

«Public Disclosure Authorized HNP DISCUSSION PAPER Public Disclosure Authorized Analyzing Changes in Health Financing Arrangements in High-Income Countries A Comprehensive Framework Approach Public Disclosure Authorized Reinhard Busse, Jonas Schreyögg, and Christian Gericke Public Disclosure Authorized February 2007 Analyzing Changes in Health Financing Arrangements in High-Income Countries A Comprehensive Framework Approach Reinhard Busse, Jonas Schreyögg, and Christian Gericke February 2007...»

«348 PEACE & CHANGE / June 2003 Land Tenure and Legal Pluralism in the Peace Process F by Jon D. Unruh OO Land tenure has proven to be one of the most vexing issues in a peace process. The disintegration of land and property rights institutions during armed conflict yet the importance of land and property to the conduct of conflict present particular dilemmas for a peace process attempting to reconfigure aspects of societal relations important to recovery. In this regard PR understanding what...»

<<  HOME   |    CONTACTS
2017 www.abstract.dislib.info - Abstracts, online materials

Materials of this site are available for review, all rights belong to their respective owners.
If you do not agree with the fact that your material is placed on this site, please, email us, we will within 1-2 business days delete him.