The Challenge of Deep Models


ion Abstraction has a central role in designing KBS. Corresponding to the levels of a KBS previously mentioned, there are di erent forms of abstraction: There is abstraction of factual knowledge. Knowledge is represented at di erent levels of abstraction supporting the inference steps to be performed at di erent levels of detail. An early example is the ABEL system [Patil 1981, Patil et al. 1982]. The importance of abstraction steps for diagnosis was also recognized early in the CADUCEUS system for internal medicine [Pople 1982] by introducing planning links associating manifestations with abstract involvement structures. The use of abstractions during diagnosis helps to focus the reasoning process. The resulting improvement of e ciency when applying abstractions on model based diagnosis has been shown by [Mozetic 1990]. But abstraction is not only of high importance to diagnosis. Planning has to look for ways to reuse solutions of one problem for a broad range of new problems. This can be achieved by building abstract plans [Tenenberg 1986]. From the viewpoint of reasoning steps, Pople's planning links form the basis of the inference step `match abstract solution'. That is the second step in the three step procedure: data abstraction { match abstract solution { re nement of solution. This procedure was found by [Clancey 1985] to be a fundamental abstract inference method when building the NEOMYCIN system. It is called `heuristic classi cation'. In the meantime several other abstract inference methods and problem solving classes have been identi ed (e.g. cover-and di erentiate, propose-and-revise, hierarchical design). This is of high importance to knowledge acquisition. Using a repertoire of standard problem solving methods it is possible to instantiate the templates of these methods [Breuker et al. 1987, Marcus 1988], if the way an expert solves a problem matches one of the abstract problem solving methods. The templates guide the knowledge acquisition process by telling what to look for. Comparable to the idea of abstract inference methods is the idea of abstract task structures: The generic task approach [Chandrasekaran 1986, 1987, 1988] has its focus on tasks like diagnosis, classi cation, or design. They are generic in the sense, that they will be instantiated to real tasks when confronted with a speci c application. A generic task de nes its function { the kind of problem it solves {, the knowledge structure and organization, and the control strategy to accomplish the function of the task. The focus is here on improvement of the knowledge acquisition process, too. Tools are provided for instantiating generic tasks: CSRL [Bylander et al. 1983], IDABLE [Sticklen 1983], DSPL [Brown and Chandrasekaran 1988]. If the application problem to solve is associated correctly with a generic task or a combination of tasks, the corresponding tool(s) will guide the knowledge acquisition process. 6 What can we expect?<lb>Let us review the methodological achievements under the aspects of the limitations of<lb>rst-generation expert systems:<lb>Robustness vs. performance: The use of structural domain knowledge and the use<lb>of principled reasoning methods based on this ne-grained domain knowledge is the<lb>way to achieve robustness when solving uncommon problems. But deep models<lb>are very slow performing knowledge structures when applying principled ways of<lb>reasoning during problem solving. Their challenge is the ability to serve as a ground<lb>for automatic construction of surface models by use of machine learning techniques<lb>analyzing cases often seen. In combination we get a fast performing problem solver<lb>when confronted with common cases, which can fall back to its robust reasoning<lb>methods on complicated situations.<lb>Explanability: The knowledge about the tasks and the inference steps the KBS is<lb>performing allows for explanations what is really going on. But not much focus has<lb>been put to explanation issues by the research community. In addition, explanations<lb>are expected to be user-tailored in the sense that they should take into account the<lb>level of experience of the user, the facts he/she already knows, and his/her intentions.<lb>This results in the need of an explicit user model. User modeling is a current area of<lb>AI research, but it will take some time until user models will guide the explanation<lb>(and reasoning) methods of expert systems in practice.<lb>Knowledge acquisition support: The support tools using the speci cation of domain<lb>models, problem solving classes and abstract task descriptions give us a chance to<lb>build KBS in a more consistent and complete way by having a speci cation of the<lb>whole problem solving process. They provide a basis to keep the system expandable<lb>and maintainable during the whole lifecycle. But (unfortunately) the knowledge en-<lb>gineer is needed more than before. At moment direct knowledge transfer from the<lb>expert to the KBS is possible only for xed domains with speci c forms of repre-<lb>sentation (e.g., OPAL [Musen 1989]). Model-based knowledge acquisition [Shadbolt<lb>and Wielinga 1990] will hopefully guide the complex knowledge acquisition process<lb>in the form of an active and directive system in the future.<lb>Of essential importance is the functional de nition of the problem solver at a concep-<lb>tual level. The speci cation of KBS at the knowledge level as instantiation of abstract<lb>knowledge structures may help us (1) to create understandable, expandable, and portable<lb>systems. We are not captured by the implementation details which often override the con-<lb>ceptual structures of the application. (2) It may give us new insights to domain models,<lb>problem solving classes and task structures, thus, enhancing our repertoire of methods.<lb>Comparing the biological versus the technical domain we recognize the in uence of<lb>medical applications not only to rst-generation expert systems, but also to model-based<lb>architectures. The comparison of these two domains raises the question: Is there a unifying<lb>7 perspective? I.e., are there completely domain-independent architectures? What speaks<lb>against is the fact that we have very di erent domains (medical vs. technical): analog sys-<lb>tems with tolerances and correcting feed-back loops vs. digital systems with 0-1 behavior<lb>and errors e ecting the whole system; partial models resulting from an incomplete under-<lb>standing of biological processes vs. complete technical models which are easy to describe<lb>formally; external (environmental) causes of errors vs. internal causes of errors (material<lb>problems); many di erent knowledge structures vs. complex reasoning processes operating<lb>on few knowledge types. This may be the reason why medical applications tend to focus on<lb>the extensive task of representing domain knowledge, whereas applications in the technical<lb>domains concentrate on nding e cient inference methods.<lb>What speaks in favor of the unifying perspective is the fact that many medical appli-<lb>cations formed the basis for more abstract methodologies, which have been used lateron<lb>in the technical domain: MYCIN ! EMYCIN, NEOMYCIN ! the inference method<lb>`heuristic classi cation', MDX/PATREC ! the generic task architecture. What seems<lb>to be identifyable are the same problem-solving methods and abstract domain concepts<lb>throughout very di erent domains. That makes abstract modeling a worthwhile thing to<lb>do. But at moment, we have only a very restricted view: (1) most systems have dealt with<lb>the problem of diagnosis, (2) the domains of application have been very restricted, and (3)<lb>we are at the early beginning of the usage of structure-based knowledge acquisition tools.<lb>In conclusion, second-generation architectures are still in the research labs. They will<lb>have a hard time to nd the way to daily practice. But they o er a lot of possibilities to<lb>overcome the limitations of currently widespread used KBS. We have a lot to expect and<lb>it seems very worthwhile to put e orts into these model-based architectures.<lb>References<lb>Bratko I., Mozetic I., Lavrac N. (1989): Kardio A Study in Deep and Qualitative Knowledge<lb>for Expert Systems, MIT Press, Cambridge, MA.<lb>Breuker J.A., Someren M.W.van, Hoog, Schreiber G., Greef, Bredeweg B., Wielemaker<lb>J., Billault J.-P., Davoodi M., Hayward S.A. (1987): Model-Driven Knowledge Acquistion:<lb>Interpretation Models, University of Amsterdam and STL Ltd., Esprit Project P1098, Memo<lb>87.<lb>Brown D.C., Chandrasekaran B. (1988): Design Problem Solving: Knowledge Structures and<lb>Control Strategies, Morgan Kaufmann, Los Altos, CA.<lb>Bylander T., Mittal S., Chandrasekaran B. (1983): CSRL: A Language for Expert Systems for<lb>Diagnosis, in Proceedings of the 8th International Joint Conference on Arti cial Intelligence,<lb>Morgan Kaufmann, Los Altos, CA.<lb>Bylander T. (1987): Some Causal Models are Deeper than Others, in Proc.Ninth Annual Confer-<lb>ence of the Cognitive Science Society, Seattle, Wash., pp.916-921 (revised version: Arti cial<lb>Intelligence in Medicine, 2(3)123-128, 1990).<lb>Chandrasekaran B. (1986): Generic Tasks in Knowledge-Based Reasoning: High-Level Building<lb>Blocks for Expert System Design, IEEE Expert, Fall, pp.23-30.<lb>8 Chandrasekaran B. (1987): Towards a Functional Architecture for Intelligence Based on Generic<lb>Information Processing Tasks, in Proceedings of the 10th International Joint Conference on<lb>Arti cial Intelligence (IJCAI-87), Morgan Kaufmann, Los Altos, CA.<lb>Chandrasekaran B., Smith J., Sticklen J. (1987): \Deep" Models and Their Relation to Diagnosis,<lb>OSU-CIS-LAIR Technical Report, Ohio State University, Columbus.<lb>Chandrasekaran B. (1988): Generic Tasks as Building Blocks for Knowledge-Based Systems: The<lb>Diagnosis and Routine Design Examples, The Knowledge Engineering Review, 3(3)183-210.<lb>Clancey W.J. (1985): Heuristic Classi cation, Arti cial Intelligence, 27(3)289-350.<lb>Davis R. (1983): Diagnosis Via Causal Reasoning: Paths of Interaction and the Locality Principle,<lb>in Proceedings of the Third National Conference on Arti cial Intelligence (AAAI-83), William<lb>Kaufmann, Los Altos, CA, pp.88-94.<lb>Davis R. (1984): Diagnostic Reasoning Based on Structure and Behavior, Special Volume on<lb>Qualitative Reasoning about Physical Systems, Arti cial Intelligence, 24(1-3)347-410.<lb>Glowinski A., O'Neil M., Fox J. (1989): Design of a Generic Information System and its Ap-<lb>plication to Primary Care, in Hunter J., et al.(eds.), Proceedings of the Second European<lb>Conference on Arti cial Intelligence in Medicine (AIME 89), Springer, Berlin, pp.221-233.<lb>Haag A. (1988): Deriving a Model of an Electrical Transformer for Qualitative Simulation with<lb>QSIM, in Fr<lb>uchtenicht H.W., et al.(eds.),Technische Expertensysteme: Wissensrepr<lb>asentation<lb>und Schlussfolgerungsverfahren, Oldenbourg, M<lb>unchen, pp.117-134.<lb>Hart P.E. (1982): Directions for AI in the Eighties, SIGART Newsletter, 79,11-16.<lb>Hayward S.A., Wielinga B.J., Breuker J.A. (1987): Structured Analysis of Knowledge, Interna-<lb>tional Journal of Man-Machine Studies, 26,487-498.<lb>Horn W. (1989): MESICAR A Medical Expert System Integrating Causal and Associative<lb>Reasoning, Special Issue on Causal Modeling, Applied Arti cial Intelligence, 3(2-3)305-336.<lb>Hunter J., Gotts N.M., Hamlet I., Kirby I. (1989): Qualitative Spatial and Temporal Reasoning<lb>in Cardiac Electrophysiology, in Hunter J., et al.(eds.), Proceedings of the Second European<lb>Conference on Arti cial Intelligence in Medicine (AIME 89), Springer, Berlin, pp.157-166.<lb>Ironi L., Lanzola G., Stefanelli M. (1989): A Qualitative Model of Iron Metabolism, in Hunter<lb>J., et al.(eds.), Proceedings of the Second European Conference on Arti cial Intelligence in<lb>Medicine (AIME 89), Springer, Berlin, pp.147-156.<lb>Keravnou E.T., Washbrook J. (1989): What is a Deep Expert System? An Analysis of the Ar-<lb>chitectural Requirements of Second-Generation Expert Systems, The Knowledge Engineering<lb>Review, 4(3)205-233.<lb>Kleer, WilliamsB.C. (1987): Diagnosing Multiple Faults, Arti cial Intelligence, 32(1)97-130.<lb>Kleer, Mackworth A.K., Reiter R. (1990): Characterizing Diagnosis, Working Notes \First<lb>International Workshop on Principles of Diagnosis", Stanford University, California, USA,<lb>July 23-25, pp.1-7.<lb>Klein D., Finin T. (1987): What's in a Deep Model? A Characterization of Knowledge Depth<lb>in Intelligent Safety Systems, in Proceedings of the 10th International Joint Conference on<lb>Arti cial Intelligence (IJCAI-87), Morgan Kaufmann, Los Altos, CA, pp.559-562.<lb>9 Kuipers B.J. (1986): Qualitative Simulation, Arti cial Intelligence, 29(3)289-338.<lb>Kunz J.C. (1983): Analysis of Physiological Behavior Using a Causal Model Based on First<lb>Principles, in Proceedings of the Third National Conference on Arti cial Intelligence (AAAI-<lb>83), William Kaufmann, Los Altos, CA, pp.225-228.<lb>Marcus S.(ed.) (1988): Automating Knowledge Acquisition for Expert Systems, Kluwer, Boston.<lb>McDermott J. (1988): A Taxonomy of Problem-Solving Methods, in Marcus S.(ed.): Automating<lb>Knowledge Acquisition for Expert Systems, Kluwer, Boston.<lb>Michie D. (1981): High-Road and Low-Road Programs, AI-Magazine, 3(1)21-22.<lb>Mozetic I. (1990): Reduction of Diagnostic Complexity through Model Abstractions, in Work-<lb>ing Notes \First International Workshop on Principles of Diagnosis", Stanford University,<lb>California, USA, July 23-25, pp.102-111.<lb>Musen M.A. (1989): Automated Generation of Model-based Knowledge-Acquisition Tools, Pitman,<lb>London.<lb>Newell A. (1982): The Knowledge Level, Arti cial Intelligence, 18(1)87-127.<lb>Patil R.S., Szolovits P., SchwartzW.B. (1981): Causal Understanding of Patient Illness in Medical<lb>Diagnosis, in Proceedings of the 7th International Joint Conference on Arti cial Intelligence,<lb>University of British Columbia, Vancouver, Canada, pp.893-899.<lb>Patil R.S., Szolovits P., Schwartz W.B. (1982): Modeling Knowledge of the Patient in Acid-Base<lb>and Electrolyte Disorders, in Szolovits P.(ed.), Arti cial Intelligence in Medicine, Westview<lb>Press, Boulder, CO.<lb>Pople H.E. (1982): Heuristic Methods for Imposing Structure in Ill-Structured Problems: The<lb>Structuring of Medical Diagnostics, in Szolovits P.(ed.), Arti cial Intelligence in Medicine,<lb>Westview Press, Boulder, CO.<lb>Reiter R. (1987): A Theory of Diagnosis from First Principles, Arti cial Intelligence, 32(1)57-96.<lb>Schreiber G., Breuker J., Bredeweg B., Wielinga B.J. (1988): Modelling in KBS Development,<lb>in Boose J.H., et al.(eds.), Proceedings of the European Knowledge Acquisition Workshop<lb>(EKAW'88), GMD-Studien Nr.143, St.Augustin, FRG.<lb>Senyk O., Patil R.S., Sonnenberg F.A. (1989): Systematic Knowledge Base Design for Medical<lb>Diagnosis, Special Issue on Causal Modeling, Applied Arti cial Intelligence, 3(2-3)249-274.<lb>Shadbolt N.R., Wielinga B. (1990): Knowledge-Based Knowledge Acquisition: The Next Gener-<lb>ation of Support Tools, in Wielinga B., et al.(eds.), Current Trends in Knowledge Acquisition<lb>(Proc.EKAW'90), IOS, Amsterdam, pp.313-338.<lb>Steels L. (1985): Second Generation Expert Systems, Future Generations Computer Systems,<lb>1(4)213-221.<lb>Steels L. (1988): The Deepening of Expert Systems, in Vichnevetsky R., et al.(eds.), Proceedings<lb>of the 12th IMACS World Congress, Paris, France, Vol.4, pp.323-326.<lb>Steels L. (1989): Diagnosis with a Function-Fault Model, Special Issue on Causal Modeling,<lb>Applied Arti cial Intelligence, 3(2-3)213-238.<lb>Steels L. (1990): Components of Expertise, AI-Magazine, 11(2)28-49.<lb>10 Sticklen J. (1983): Manual for IDABLE, OSU-CIS-LAIR Technical Report, Ohio State University,<lb>Columbus.<lb>Sticklen J., Chandrasekaran B. (1989): Integrating Classi cation-Based Compiled Level Reason-<lb>ing with Function-Based Deep Level Reasoning, Special Issue on Causal Modeling, Applied<lb>Arti cial Intelligence, 3(2-3)275-304.<lb>Sticklen J., Chandrasekaran B., Bond W.E. (1989): Applying a Functional Approach for Model<lb>Based Reasoning, in Proceedings of the IJCAI-89 Workshop on Model-Based Reasoning, De-<lb>troit, Mich., pp.165-176.<lb>Tenenberg J. (1986): Planning with Abstraction, in Proceedings of the Fifth National Conference<lb>on Arti cial Intelligence (AAAI-86), Morgan Kaufmann, Los Altos, CA.<lb>Washbrook J., Keravnou E.T. (1990): Making Deepness Explicit, Arti cial Intelligence in Medi-<lb>cine, 2(3)129-134.<lb>Weiss S.M., Kulikowski C.A., Amarel S. (1978): A Model-Based Method for Computer-Aided<lb>Medical Decision-Making, Arti cial Intelligence, 11,145-172.<lb>11

Extracted Key Phrases

Cite this paper

@inproceedings{Horn1990TheCO, title={The Challenge of Deep Models}, author={Werner Horn}, year={1990} }