WWW.ABSTRACT.DISLIB.INFO
FREE ELECTRONIC LIBRARY - Abstracts, online materials
 
<< HOME
CONTACTS



Pages:   || 2 | 3 | 4 | 5 |   ...   | 8 |

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

-- [ Page 1 ] --

The Use of Explicit Goals for Knowledge to Guide

Inference and Learning

Ashwin Ram Lawrence Hunter

College of Computing National Library of Medicine

Georgia Institute of Technology Building 38A, Mail Stop 54

Atlanta, Georgia 30332-0280 Bethesda, MD 20894

(404) 853-9372 (301) 496-9300

E-mail: ashwin@cc.gatech.edu E-mail: hunter@nlm.nih.gov

Abstract Combinatorial explosion of inferences has always been a central problem in arti cial intelligence.

Although the inferences that can be drawn from a reasoner's knowledge and from available inputs is very large (potentially in nite), the inferential resources available to any reasoning system are limited.

With limited inferential capacity and very many potential inferences, reasoners must somehow control the process of inference.

Not all inferences are equally useful to a given reasoning system. Any reasoning system that has goals (or any form of a utility function) and acts based on its beliefs indirectly assigns utility to its beliefs. Given limits on the process of inference, and variation in the utility of inferences, it is clear that a reasoner ought to draw the inferences that will be most valuable to it.

This paper presents an approach to this problem that makes the utility of a (potential) belief an explicit part of the inference process. The method is to generate explicit desires for knowledge. The question of focus of attention is thereby transformed into two related problems: How can explicit desires for knowledge be used to control inference and facilitate resource-constrained goal pursuit in general?

and, Where do these desires for knowledge come from? We present a theory of knowledge goals, or desires for knowledge, and their use in the processes of understanding and learning. The theory is illustrated using two case studies, a natural language understanding program that learns by reading novel or unusual newspaper stories, and a di erential diagnosis program that improves its accuracy with experience.

Journal of Applied Intelligence, 2(1):47{73, 1992.

1 The focus of attention problem Combinatorial explosion of inferences has always been a central problem in arti cial intelligence. Although the inferences that can be drawn from a reasoner's knowledge and from available inputs is very large (potentially in nite), the inferential resources available to any reasoning system are limited. In general, reasoning systems simply cannot draw all justi ed inferences. With limited inferential capacity and very many potential inferences, reasoners must somehow control the process of inference.

Not all inferences are equally useful to a given reasoning system. Any reasoning system that has goals (or any form of a utility function), and acts based on its beliefs, indirectly assigns utility to its beliefs. That is, some beliefs have a causal role in actions that, in turn, lead to good outcomes. Beliefs arise, at least in part, by inference; therefore some inferences lead to better outcomes than others. Given limits on the process of inference, and variation in the utility of inferences, it is clear that a reasoner ought to draw the inferences that will be most valuable to it. How can that be done?

Several methods of controlling inference have been proposed. Perhaps the simplest is constrained forward chaining: making as many inferences as possible within the resource constraints. For example, MARGIE Rieger, 1975] made all the justi ed inferences that required the chaining of no more than 5 of 17 rules. The amount of inference scales with the number of inference rules to the power of the length of possible chains, so in practical circumstances only a small percentage of justi ed inferences can be drawn with this method.

Empirically, many of the inferences generated this way are useless, and many useful inferences are missed because their derivations are too long. Other systems rely on backward chaining to make only the inferences that might lead to a speci ed outcome. Unfortunately, many valuable inferences (even quite simple ones) are overlooked by this method, since surprises are impossible; the system can only infer what it's already looking for. Still another method tries to use probability measures to draw the most likely inferences. However, some relatively unlikely inferences can be extremely valuable. Reasoners may be explicitly interested in identifying unlikely events that have signi cant consequences. Simple counterexamples demonstrate the inadequacy of each of these approaches in the general case.

Our approach to this problem has been to make the utility of a (potential) belief an explicit part of the inference process. The method is to generate explicit desires for knowledge. The question of focus of attention is thereby transformed into two related problems: How can explicit desires for knowledge be used to control inference and facilitate resource-constrained goal pursuit in general? and, Where do these desires for knowledge come from? To address these questions, we must consider the uses of knowledge and inference, and how to assess the value of knowledge in order to control inference. To illustrate our approach, we present two implementations of these ideas: IVY and AQUA. AQUA is a natural language understanding program that learns by reading unusual newspaper stories, and IVY is a medical diagnosis program that improves its accuracy with experience. Both of these programs use knowledge goals to control their processing.





1.1 Inference and desires for knowledge Inference plays a wide variety of roles in reasoning systems. For example, understanding natural language texts requires more than just identifying the literal meanings of words. Human understanders \read between the lines," making a large number of inferences not directly contained in a text. For example, if told that \John took some aspirin," most people would infer that he was in some pain, that he drank some water with them, that he swallowed them (rather than stole them), and so forth. As early natural language researchers

discovered, the number of inferences that can be drawn from even a simple sentence is potentially in nite:

There were fewer aspirin in the bottle, there was a bottle that the aspirin came in, the bottle was the size, shape and color of the usual over the counter medication bottle, the aspirin were the size of pills, were round and white, John felt better about 20 minutes later, etc., etc. It is not hard to nd a story where being able to infer any of these facts is important to understanding the story, yet it is computationally intractable to make all the inferences that can be drawn in any given situation. Somehow, people are able to manage the many possible inferences, generally without missing important details or taking a long time to understand.

A similar problem arises in abductive (e.g., diagnostic) systems. Abduction, the construction of causal explanations, is often viewed as inference to the \best" explanation. However, the de nition of \best" is, as before, dependent on the goals of the reasoner in forming the explanation and not just on the correctness of the causal chain underlying the explanation Ram and Leake, 1991]. In situations where there is not just a single correct explanation, the best explanation must address the reason that the explanation was required in the rst place. For example, if the purpose of an explanation is to avoid repetition of a failure, the explanation should be generalizable to similar future situations. Often, the operational de nition of \best" explanation includes some component of its utility to the reasoner. The point here is that the value of abduced knowledge should play a role in how an abductive process works.

Another instance of the inference control problem arises in the design of machine learning systems in general. It can be formally demonstrated that far more inferences are licensed by induction over a set of experiences than can be distinguished among using those experiences Dietterich, 1989]. In general, an inductive system must have a method for preferring some inferences over others. Existing machine learning methods have done this by applying inductive biases (e.g., Utgo, 1986]), or by a priori limitations on the structure of the inferences they can make, through, for example, the use of decision trees or neural networks.

These approaches can be considered syntactic, in that they constrain the form of the hypotheses considered, rather than their content.

The problem arises in noninductive learning systems as well. For example, the questions of \when to generalize" and \how far to generalize" are among the central issues in explanation-based learning. Again, most of the approaches to this problem thus far involve syntactic solutions. An exception is Minton's PRODIGY system Minton, 1988], which evaluated each explanation's e ect on the average reasoning process before integrating it into permanent memory. Although Minton did use the utility of an inference directly in his machine learning system, the utility of the inference did not play a role in the generation of the inference, only in deciding whether to store it.

If inference for learning must be constrained, it should be directed towards achieving (or at least facilitating) the goals of the performance system that the learner is part of. When a reasoner encounters di culties during understanding, planning, or whatever else its task is, it should be able to remember the nature of these di culties, and learn in order to become better at the tasks that it is trying to perform.

This characterization of knowledge that it would be useful to have provides a valuable tool for determining the utility of knowledge and inference later on.

We propose that the method of restricting potentially inferrable hypotheses should be content-based.

Explicit characterizations of desirable knowledge or required knowledge provide a principled method for restricting the realm of experience and background knowledge considered in inference, and thereby the size of the hypothesis space that must be considered. Having goals specifying what (kind of) knowledge is desirable provides a signi cant advantage for systems trying to learn from very complex experience.

In fact, we can carry this idea one step further: Not only can explicit goals about knowledge help control inference, they can be used to direct action intended to accomplish those goals. Rather than passively waiting for useful information to show up, a system can actively pursue the knowledge it desires, using speci c learning plans or instantiations of general learning strategies. In order to actively plan to learn, as well as for control of inference in general, a reasoning system needs to represent and reason about its own desires for knowledge, and consider them actively in order to make decisions during the inference process.

This paper presents a theory of inference control for understanding and learning that is based on the notion of knowledge goals, a reasoner's speci c desires to acquire and organize useful beliefs. This theory is broadly applicable to automated reasoning systems that improve with experience. A knowledge goal represents the need to ll in gaps in the reasoner's knowledge base that are detected when a piece of information required for a task turns out to be missing, incorrect or otherwise problematic. Our theory addresses the representation of knowledge goals, methods for introspective reasoning about the reasoner's own knowledge to generate goals, heuristics for inference control and hypothesis evaluation using goal-based focus of attention criteria, as well as algorithms for learning through the active pursuit of knowledge goals.

In addition, this theory suggests a method for controlling the timing of inference. Programs don't always know everything they need to know. A program that learns has the additional problem of having what it knows change over time. An inference that may be di cult to make at one time may be much easier to make later, when additional information is available. Explicit knowledge goals make possible opportunistic learning over an extended period of time. A learning program intended to run inde nitely can use knowledge goals to make decisions about when to learn, as well as what to learn. Both of the programs described below manage a list of pending knowledge goals, and notice opportunities to achieve them during other processing, long after the goals have been generated (and after attempting to address them at that time). To our knowledge, this is a unique ability in machine learning programs, and it is mediated by explicit management of knowledge goals.

1.2 Do people have goals for knowledge?

Our theory is based both in a theoretical analysis of the constraints on inference in practical AI systems, and in empirical psychological evidence. People quite clearly have what psychologists often call \goal orientations," which have a signi cant e ect on the inferences that people draw from their experiences. There is a large body of psychological research on goal direction in focus of attention, particularly from social psychology. Zukier's 1986] review concludes: \Experimental studies have clearly demonstrated that a person will structure and process information quite di erently, depending on the future use he or she intends to make of it. Information integration clearly is preceded by future-oriented decision-making processes, which guide data selection and the choice of an appropriate strategy or mode from among the several that are available," p. 495].

Ho man, et al 1981] demonstrate that di erent goal orientations (e.g., \form an impression of a person in the following story" or \remember as much as you can from the following story") may in uence not only the use of di erent representations, but also the selection among di erent kinds of processing. Although the goal orientations tested in that work are quite abstract, they signi cantly constrain the space of hypotheses consistent with the experimental materials. Srull and Wyer's 1986] results, although divergent in important respects from those of Ho man, et al, also provide evidence that di erent goal orientations have a strong e ect on learning.

In addition to the empirical psychological ndings, consideration of the di erences between how existing computer programs parse newspaper stories (e.g., FRUMP DeJong, 1979]) and how people read them

supports this approach. These di erences include:



Pages:   || 2 | 3 | 4 | 5 |   ...   | 8 |


Similar works:

«FREQUENTLY ASKED QUESTIONS Basic Questions: What is included in the LV Series Kit Home? What is an LV Series Home? Is the LV modular or stick-built? What is the minimum space required to build an LV Series Home? Does the LV Series Home meet residential building codes everywhere? Do the materials used in your manufacturing process pose potential health problems? Can you please send me a catalog or brochure? Sales/Pricing Questions: What is the price and what is included in the LV Series Home?...»

«Acta Scientiae et Intellectus ISSN: 2410-9738 PAINLESS “HUNGER-DEATH” IS A MYTH PhD, MSc V.J.T. van Ginneken Blue-Green technologies Foundation; www.bluegreentechnologies.nl, THE NETHERLANDS E-mails: vvanginneken@hotmail.com, vincent@bluegreentechnologies.nl ABSTRACT Since at present A.D. 2015 around one billion people are exposed to hunger ( 200 kcal/caput/day) -among them around 7 million children (aged 5 years) and around 4 million neonatal deathswe will discuss the important question...»

«Investigation of the Brain Regions Involved in Improvisation Richard Falco1, Frederick Bianchi1, and Karl G. Helmer2,3, Department of Humanities and Arts, Worcester Polytechnic Institute Athinoula A. Martinos Center of Biomedical Imaging, Massachusetts General Hospital Department of Radiology, Harvard Medical School. Background and Significance Improvisation in music is the generation of a new melody (with or without the accompanying harmony) in response to a written melody or harmonic...»

«Package leaflet: Information for the user Nicabate Menthol Mint 2 mg and 4mg medicated chewing gum Nicotine Read all of this leaflet carefully before you start using this medicine because it contains important information for you. Always use this medicine exactly as described in this leaflet or as your healthcare professional has told you.Keep this leaflet. You may need to read it again.Ask a healthcare professional e.g. doctor, nurse, smoking cessation advisor or pharmacist if you need more...»

«Nos. 14-1418, 14-1453, 14-1505, 15-35, 15-105, 15-119 & 15-191 ================================================================ In The Supreme Court of the United States -MOST REVEREND DAVID A. ZUBIK, et al., Petitioners, v. SYLVIA BURWELL, Secretary of Health and Human Services, et al., Respondents.-On Writs of Certiorari To The United States Courts Of Appeals For The Third, Fifth, Tenth And District Of Columbia Circuits -AMICUS CURIAE BRIEF OF ASSOCIATION OF AMERICAN PHYSICIANS &...»

«Touching Trauma The use of touch for supporting resilience and development of the capacity for self-regulation Kathy L. Kain, M.A. Abstract Resiliency and self-regulation are primary elements that contribute to recovery from traumatic stress. Presented here are four touch-based methods for enhancing your clients’ capacity to stay somatically present, regulate physiological responses, and minimize the potential for being overwhelmed by strong physiological sensations. Such somatic skills on...»

«TOXICOLOGICAL SUMMARY AND SUGGESTED ACTION LEVELS TO REDUCE POTENTIAL ADVERSE HEALTH EFFECTS OF SIX CYANOTOXINS May 2012 Office of Environmental Health Hazard Assessment California Environmental Protection Agency Blank Page May 2012 TOXICOLOGICAL SUMMARY AND SUGGESTED ACTION LEVELS TO REDUCE POTENTIAL ADVERSE HEALTH EFFECTS OF SIX CYANOTOXINS FINAL REPORT May 2012 Office of Environmental Health Hazard Assessment California Environmental Protection Agency 1001 I Street, 12th Floor P.O. Box 4010...»

«IOSR Journal of Dental and Medical Sciences (IOSR-JDMS) e-ISSN: 2279-0853, p-ISSN: 2279-0861.Volume 15, Issue 1 Ver. II (Jan. 2016), PP 10-17 www.iosrjournals.org Use of Information technology among Physiology postgraduates in Tamilnadu and Pondicherry Dr.G.Sandeep*, Dr.Semmal1, Dr.Sheela Ravinder.S2 *Physiology post graduate, Sri Ramachandra University, Tamilnadu, India Associate professor in Physiology, Satya Sai Medical College,Tamilnadu Associate professor in Physiology, Sri Ramachandra...»

«ADMISSION TEST FOR THE DEGREE COURSE IN MEDICINE AND SURGERY Academic Year 2013/2014 Thinking Skills Ever since Uranus was discovered, astronomers have thought there might be more planets in the Solar System. Because of small deviations in the orbits of Uranus and Neptune deviations which would occur if another planet existed some astronomers think there must be an undiscovered planet Planet X. But these deviations cannot tell us whether Planet X exists, because they would occur if the orbits...»

«THE UNIVERSITY OF TEXAS AT TYLER FACULTY SCHOLARSHIP JANUARY – DECEMBER 2014 INGENUITY CENTER The Ingenuity Center is an Interdisciplinary Research and Development center that brings together faculty, students, and staff from the College of Education and Psychology, College of Engineering, the College of Nursing and Health FROM THE PRESIDENT Sciences and independent researchers in an effort to improve STEM Education across Texas. In 2014, the Ingenuity Center secured grants and contracts in...»

«QUT Digital Repository: http://eprints.qut.edu.au/ Beesley, Vanessa L. and Eakin, Elizabeth G. and Steginga, Suzanne and Aitken, Joanne F. and Dunn, Jeff and Battistutta, Diana (2008) Unmet needs of gynaecological cancer survivors: implications for developing community support services. Psycho-Oncology, 17(4). pp. 392-400. © Copyright 2008 John Wiley & Sons FULL TITLE: Unmet needs of gynaecological cancer survivors: implications for developing community support services SHORT TITLE: Needs of...»

«For Members of the News Media Childhood Traumatic Grief Educational Materials From the National Child Traumatic Stress Network Childhood Traumatic Grief Task Force Educational Materials Subcommittee This project was funded by the Substance Abuse and Mental Health Services Administration, U.S. Department of Health and Human Services Childhood Traumatic Grief Educational Materials For Members of the News Media From the National Child Traumatic Stress Network Childhood Traumatic Grief Task Force...»





 
<<  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.