PII: 0957-4174(96)00018-8 Pergamon Expert Systems With Applications, Vol. 10, No. 3/4, pp. 393--401, 1996 Copyright © 1996 Elsevier Science Lid Printed in Great Britain. All rights reserved 0957-4174/96 $15.00+0.00 S0957-4174(96)00018-8 Interactive Induction of Expert Knowledge BINGCHIANG JENG, TING-PENG LIANG AND MINYANG HONG Department of Information Management, National Sun Yat-Sen University, Kaohsiung, Taiwan 80424, Republic of China Abstract--The process o f extracting, structuring and organizing elicited knowledge (called knowledge acquisition) is a bottleneck in developing knowledge-based systems. A manual approach that elicits domain knowledge by interviewing human experts typically has problems, because the experts are often unable to articulate their reasoning rules. An automatic approach that induces knowledge from a set o f training cases also suffers from the unavailability o f sufficient training cases. We present an integrated approach that combines the strengths o f both methods to compensate f o r their weaknesses. In this approach, human experts are responsible f o r solving problems, whereas an inductive learning algorithm is responsible f o r reasoning and consistency checking. Copyright © 1996 Elsevier Science Ltd 1. I N T R O D U C T I O N K N O W L E D G E ACQUISITION is a process o f extracting, structuring and organizing elicited knowledge from domain experts or other knowledge sources and convert- ing the knowledge into rules or other forms o f representation accessible with a computer program [16,25]. As the power o f an expert system comes from the knowledge that it possesses [13], the acquisition o f knowledge is a major stage in the development of an experts system. The transfer of expertise from various sources to a knowledge base is difficult and challenging. A manual process that encodes domain knowledge by interviewing human experts suffers from several difficulties [31,40]. For example, human experts are typically not also expert in articulating their rules of reasoning because they are not explicitly aware o f the structure o f their knowledge. A paradox o f expertise [20] claims that the better one is an expert, the worst one is to tell the details. Knowledge acquired during interview may be unreliable, as verbal reports and mental behavior are not necessarily corre- lated [2]. Incongruities may occur among what an expert says that he does, what he actually does and what he should have done [17]. Communication between the knowledge engineer and the domain expert is another difficult problem. The knowledge engineer may know little about the problem domain, and may not understand clearly the jargon used by the domain expert. On the other hand, the domain expert may not understand the process o f knowledge acquisition and what the knowledge engineer needs. Even the cultural differences between the two parties o f knowledge acquisition may undermine the development o f a knowledge-based system. When historical examples are available, an automatic process can be used to induce rules from fragments o f knowledge [35,36]. The unique power o f this approach is the "automatic" learning capability o f its inductive algorithm. Many successful applications are reported, such as disease diagnoses [42,23,29], business applica- tions [5,8,39,22] and others [5,10,26]. A major constraint o f automatic learning is that it requires many training cases from which to induce knowledge. The quality o f the training result depends on the quality o f the data. Another drawback is that induction may generate knowledge that uses paths of reasoning different from those used by the expert. This may make human experts reluctant to accept the induced knowledge base regardless o f its performance. Given that each approach alone has limitations and their respective strengths and weaknesses seem com- plementary, we seek to integrate the two approaches to gather the advantages and to avoid the deficiencies. The idea is that the weaknesses o f one approach be compensated by the strengths o f the other. For example, the task o f articulating a reasoning process for a human expert can be left to an inductive method to solve, or the restriction o f inductive learning on training examples can be given to an experienced human expert to provide help. This approach is useful as human experts are good at solving problems whereas inductive learning algorithms are good at extracting rules o f reasoning. Previous research that aimed at a similar goal exists. For example, Parsaye [31] proposed an approach that combines interactive knowledge acquisition with rule induction. Based on the theory o f repertory grids [3,21], the system can interactively interview an expert by 393 394 Bingchiang Jeng et aL asking questions to capture the expert's knowledge. The captured knowledge is then generalized using rule induction. Buntine and Stirring [7] presented another approach that allows a human expert to interact with a set o f automatically induced rules so that each other's knowledge is cross-checked. Recently Evans and Fisher [12] reported a similar approach that is successfully applied to a problem domain where the experts have only weak causal knowledge. A limitation o f Buntine and Stirling's approach is that it still depends on a set o f pre-existing training examples. In some cases, these examples are hard to come by. Furthermore, the way an expert interacts with the preliminary rules is informal. The approach presented in this paper extends their ideas to allow interactive knowledge acquisition from (nearly) scratch. A modified algorithm for inductive learning is designed so that an expert communicates with the system in an interrogative style during the inductive process. In addition to inducing a decision tree from existing training cases, the system also identifies near- miss cases falling on the classification borders for clarification and verification by human experts. These cases, after classification, become new training cases for accurate learning by the modified inductive algorithm. This way, the expert's knowledge is accumulated incre- mentally and the learning process can be repeated until the induced knowledge is satisfied by the human expert. The remainder o f the paper is organized as follows. The process of automatic knowledge acquisition and its related techniques is reviewed in Section 2. An inter- active process for inductive learning that shows how to elicit knowledge from a domain expert is described in Section 3. Experimental results from the evaluation of the performance o f the proposed approach is presented in Section 4. Section 5 summarizes the research. 2. A U T O M A T I N G K N O W L E D G E A C Q U I S I T I O N There are many different methods for knowledge acquisition, which can be classified as: manual, semi- automatic and automatic [11,25]. Eliciting knowledge by manual methods is highly labor-intensive. Methods belonging to this category include structured/unstruc- tured interviews, analysis o f protocol, observations and so forth [40]. Semi-automatic methods are primarily designed to support either the expert or the knowledge engineer to perform the necessary task more effectively. Tools to elicit domain knowledge include ETS, KRITON, AQUI- NAS, MORE, M O L E etc. [4,24]. Others include TEIRESIAS [17] that assists the knowledge engineer to update a knowledge base, and KADS [14] that provides a collection o f tools to support a knowledge engineer to extract, to structure, to analyze and to document expert knowledge. Although semi-automatic methods expedite the work o f knowledge engineers and/or domain experts, they still share problems o f manual methods: the acquired knowl- edge is difficult to validate; correlation between verbal reports and mental behavior may be weak. In addition, Michie [28] stated that knowledge o f certain types could not be elicited using manual methods, simply because the domain was so large and complicated that the expert would be unable to explain how it operates. Automatic methods that use machine-learning tech- niques to extract knowledge require less or no participation by either knowledge engineers or domain experts. Therefore, they do not have the difficulties associated with human experts. Currently, the most popular automated method is rule induction. Induction is a process o f general inference from particular instances. Rule induction (or inductive learn- ing) refers to a concept learning process by which a set o f rules is created from training cases to explain or to predict a problem-solving behavior. When the induced structure o f knowledge is represented in the form o f a decision tree, it is also called (decision) tree induction [27,35]. A decision tree is considered as a set o f rules in a compact form; it can be transformed into rules easily. Early work on rule induction is traced to 1966 when Hunt, Martin and Stone developed a method for induction. The method was later implemented and expanded b y Paterson and Niblett [32] to create ACLS (A Concept-Learning System) and by Quinlan [34,35] to develop the popular ID3. In the following, we shall use the ID3 algorithm to illustrate how rule induction works. Input to ID3 is a collection o f training cases. Each is described by a set o f attributes associated with a class name (or outcome). An attribute can be either categorical (e.g. color) or numerical (e.g. age). A numerical attribute can adopt discrete or continuous values. The basic procedure o f ID3 applies a divide-and-conquer approach to partition recursively the data set, based on a test on selected attribute values, into mutually exclusive subsets. The procedure continues until each subset contains cases o f the same class (to avoid over-fitting the training data, termination conditions may be defined) or no attribute is available for further decomposition. If S= {(ai~ . . . . . ain; ci)laij~A j, ci~C)} is a set o f training cases, where Aj denotes the domain o f attribute j and C for the class, the algorithm is shown in Fig. 1. An example shown in Fig. 2 is the decision tree created by ID3 from financial data to analyze the risk of bankruptcy. A new case is classified by examining which leaf it reaches when traveling down through the tree. The traversed path is determined b y a sequence of tests to the new case (i.e. a branch is selected depending on the outcome o f a test that the new case generates). When it reaches a leaf in this way, it is considered similar to other existing cases contained in the leaf as they have all passed the same tests. Hence, the outcome o f the new case is the same as its neighbors and is assigned with the I n t e r a c t i v e I n d u c t i o n o f E x p e r t K n o w l e d g e 3 9 5 Procedure I n d u c t i o n (S: TrainSet, T: Tree) Begin If S contains cases of the same class Then label T with the class name and exit Else Begin For each numerical attribute A i, Do Find a value v/to decompose the training set into two subsets, Calculate the entropy of the decomposition, Choose the decomposition whose entropy value is the smallest; For each categorical attribute, Decompose S by its classes and Calculate its entropy; Partition S into two or more mutually exclusive subsets: S i, i = 1,..., k, based on the attribute A s whose entropy value is the smallest after decomposition. Create a node T i for each subset S i, and link it to the parent node as its child. For i = 1 to k, call I n d u c t i o n (S i, T i) End End, FIGURE 1. T h e ID3 algorithm. class n a m e labeled at the leaf. F o r e x a m p l e the prediction o f b a n k r u p t c y for case (0, 0, 0, 0.02, 0.05, 0.6, 0.04) is YES, as it follows the path at the far left d o w n to a leaf labeled " Y E S " . 3. I N T E R A C T I V E I N D U C T I O N T h e fundamental basis for methods o f decision tree induction (or rule induction) is that t w o cases with similar features are classified into the same class. The success o f such an a p p r o a c h relies on sufficient training cases to c o v e r every aspect o f the p r o b l e m d o m a i n w i t h o u t inconsistency. Otherwise, the i n d u c e d k n o w - ledge m a y be unreliable. In practical applications, however, this a s s u m p t i o n is c o m m o n l y violated. T h e sources o f training cases are, in general, f r o m a domain expert o r f r o m historical data in a maintained database. Difficulties arise typically because the domain expert can provide only a few selected examples, whereas historical data, if they exist, m a y be obsolete o r contain errors and missing values. A further difficulty with inductive learning is in its explanation capability. Explaining the reasoning process b y which a conclusion is r e a c h e d is a key requirement for an expert system. Rules i n d u c e d f r o m the training set may, however, differ f r o m those used b y the expert. This makes its explanation s o m e t i m e s less acceptable to h u m a n beings because the u n d e r l y i n g process o f reason- ing used b y the expert s y s t e m m a y be incomprehensible to users. It is thus useful to have interactions between h u m a n experts and inductive learning algorithms. This allows h u m a n experts to provide useful knowledge, such as hand-crafted tutorial examples, rules o f thumb, general hints and problem-solving strategies, to help an inductive algorithm. The inductive algorithm presented b e l o w is an example that supports interaction with experts during its learning process. After a k n o w l e d g e structure is induced f r o m its training cases, the a l g o r i t h m identifies cases that cannot be correctly classified. T h e s e cases are b r o u g h t to the expert in the f o r m o f questions to be solved. O n c e they are solved, they b e c o m e n e w training cases to the inductive algorithm for further learning. In this way, expert k n o w l e d g e is incrementally elicited and incorpor- ated into the induced k n o w l e d g e structure. A typical scenario o f the interactive inductive process is as follows. Initially, the k n o w l e d g e e n g i n e e r collects k n o w l e d g e f r o m all possible sources. Different forms o f k n o w l e d g e are stored differently. Rules (mostly f r o m domain experts) are stored in the rule set and tutorial cases (mostly f r o m the historical database) are stored in the training set. Then, the algorithm tests any inconsistencies L e a f 1 < ' ~ Leaf4 Leaf2 Leaf3 L e a f 5 L e a f 6 B a n k r u p t c y : Y e s , N o C I : C o n s i s t e n c y O p i n i o n C 2 : Subject to O p i n i o n C 3 : G o i n g C o n c e r n QI : Net Income/Total Assets Q 2 : Current Assets/Total Assets ( ~ : Current R a t i o : Cash/Total Assets FIGURE 2. A n e x a m p l e of a decision tree for bankruptcy prediction. 396 Bingchiang Jeng et al. Y Xc a l a a a x < . x / • • • " b ~ - ~ Y < ~ / ~Y>Ye a a • b Sb Sc Xc a a b b Sc Y c Sb b X (a) A decision tree (b) Partition of feature space FIGURE 3. S p a c e partition by a decision tree. in the knowledge structure represented by these two sets. As both sets are incomplete at the beginning, contra- dictory cases (i.e. cases classified differently by those two knowledge sources) are identified. These cases are then presented to the domain expert for further review. Prompting experts to review contradictory cases has two purposes. First, certain knowledge ignored by the expert may be use to provoke him to describe the inference rules in more detail. Then the contradictory cases, after correction, become new examples to the inductive algorithm for further learning. Revision o f the two sets triggers the cross-validation process again and the above procedure is repeated. A sketch of this approach is shown in Fig. 5. An important feature of the above algorithm is its ability to detect inconsistencies during the process o f interactive induction. This is important when two different sources o f knowledge are merged because Sa X inconsistencies may exist. In the following we shall present an algorithm for detecting inconsistencies. `% Sb G i v e n b o rd er • • m e m o . • a e o Y Con-ect border 3.1. S t r a t e g y f o r B o r d e r V a l i d a t i o n A decision tree created from training cases can be represented as a partitioned feature space. The induction o f decision trees (or rules) in some sense is a process to maximize the internal similarity within the partitioned subspaces. Each leaf o f the tree corresponds to a subspace, and each case in the problem domain corre- sponds to a point in it. For example, a decision tree shown in Fig. 3 partitions a two-dimensional feature space with attribute (X, Y) into three subspaces. Sub- spaces S, and St represent class A and subspace Sb represents class B. A new case is evaluated based on the subspace into which it falls (e.g. a case with X - 3 and Y~<3 and Y ~ > - 3 Then Class = 1 Else Class=0; The concept learned after the first iteration o f the induction is given in Fig. 6(b), from which the contra- dictory cases identified are shown in Fig. 6(c). The final concept induced after three iterations o f the learning process is given in Fig. 6(d). One can see that the result approaches the actual circle pattern. In addition to the nine given training cases, the system generates a total o f 20 questions for the human expert to answer in the learning process. Figure 7 shows how other geometric patterns can be learned in experiments. They all achieve satisfactory results. The learned patterns are near the actual patterns. Interactive Induction of Expert Knowledge 399 y X X (a) Initial Training Cases ~ . ~ l a s s 0 (b) First Inductive Result Y ~ . . . . , , o . . - _x :::Cia s i : : : . : . ; . ; . ; . : ( c ) Identified Conflicting (d) Final Inductive Result Cases FIGURE 6. A learning process for the circle experiment. The only limitation is that the nature o f the orthogonal partition with ID3 does not allow a real curvilinear draw. 4.2. Learning Blackjack Games Another experiment shows how a system can learn the rules that a human expert uses in playing blackjack. The game is played between a dealer and players. Each player receives two reversed cards, while the dealer receives one reversed and one obverse cards. An ace can count as either 1 point or 11 points. During the game, a player m a y request for more cards (to "hit") or to stay with the current hand (to "stand"). A player wins if the total points o f cards in hand is below 21 points but greater than the dealer's. A human expert decides whether to hit or to stand based on his current count in hand, with or without an ace, and the dealer's obverse card. These three attributes determine a decision. The expert's rules can be drawn in a decision diagram as shown in Fig. 8(a). Initially five randomly chosen cases were given for training. The rules learned after three iterations are shown in Fig. 8(b), which are pretty close to the actual rules used b y the expert. 5. RELATED WORK The subject discussed in this article can also be seen as a problem o f concept learning in the machine learning field. In particular, the way a domain expert interacts with a learning system, can be formulated as a model o f learning with membership queries [1], in which an instructor or oracle exists to answer any queries placed by a learner. In this model, the learner has control o v e r what part o f information it receives from a problem domain next. Although our work starts initially from a different perspective, its result happens to be in accordance with most recent work in machine learning. Cohn et al. [9] presented an approach to learn a concept by generating queries from a region o f uncertainty, an area in the domain where misclassification is still possible. An interesting coincidence in their neural network's imple- mentation is that it uses two networks S and G to identify an uncertain case, when outcomes o f the case are inconsistent between them. They demonstrated in several domains that the approach gives better learning results for a fixed number o f training cases than simply learning from examples alone. This is encouraging since the rationale o f identifying regions o f uncertainty is similar to our identification o f contradictory cases. Another learning model similar to the above is the on- line (or incremental) learning model in which the learner answers a sequence o f yes/no questions with immediate feedback provided after each question. A variant o f the on-line learning model, called self-directed learning, has been recently proposed b y Goldman and Sloan [15] which allows the learner to select the presentation order for the instance. Thus it can be seen as a variation o f learning with membership queries in which the learner is only "charged" for queries whose outcomes are unpre- Y . : . : . . . : - : . --i.',';:i:i:i;:i:::i.;.,~'~" ' • .~ ~ (a) ~ Class 0 x ~,:'J ..v.k-.~ t ; [ . " x . ' . ~.::;: !!!i (b) FIGURE 7. L e a r n i n g other geometric patterns. Class 0 (c) 400 Bingchiang Jeng et al. P l a y e r ' s Current Count Player's Current Count 21 2 0 19 18 17 I 6 15 14 13 12 11 10 D e a l e r ' s o b v e r s e c a r d 3 4 5 6 7 g 9 1 0 A - - S t a n d . . . . - - - H i t . . . . 21 2 0 19 15 17 16 15 14 13 12 11 10 2 3 4 5 6 7 8 9 1 0 A - - - S t a n d _ - - With Ace Without Ace ( a ) E x p e r t s ' r u l e s to p l a y b l a c k j a c k Dealer's obverse card 3 4 5 6 7 8 9 1 0 A 3 4 5 6 7 g 9 1 0 A 21 21 2o - - S t a n d - - - ~ 2o _ _ _ Stand 19 19 18 18 17 17 16 16 15 15 14 - H i t - - - H i t - - 14 13 13 12 12 11 11 1o lo Hit W i t h A c e W i t h o u t A c e ( b ) R u l e s l e a r n e d a c c o r d i n g t o t h e i n d u c t i v e a l g o r i t h m FIGURE 8. Learning the blackjack playing rules. • S t a n d . . _ H i t _ dictable. Theoretical results given by Goldman and Sloan show that the performance of self-directed learning is the best among all other commonly studied on-line and query learning models. Our process of learning by detecting contradictory cases near the classification border is also in accordance with the main results discovered from an autonomous exploratory learning system (i.e. without an external teacher). Winston [43] first drew attention to the critical role of near-miss training examples. Recently the Protos system, developed by Porter et al. [33], also used near- miss cases to learn an examplar difference before matching a new case to an existing examplar. 6. D I S C U S S I O N A N D C O N C L U S I O N S In this article, we have presented an algorithm for interactive induction and evaluated the performance of the implemented system in learning goemetric patterns and blackjack game strategies. The algorithm is simple but efficient. The contribution of this work is two-fold. First, it suggests an effective way o f integrating manual knowl- edge acquisition and inductive learning to achieve a more accurate knowledge base. Second, it provides a new way of creating near-miss training examples and learning from these examples. This allows critical knowledge to be learned without a large number o f training cases. Further research following this includes replacing ID3 with other learning methods, testing in other domains and exploring the handling o f certainty factors in interactive induction. The current approach can also be extended to acquire knowledge from multiple experts. The main difficulty for knowledge acquisition from multiple experts is how to combine their expertise and find out any inconsistencies systematically. Due to different subjective opinions, two competitive domain experts may disagree with each other occasionally. Our approach should be helpful in identifying conflicting rules or cases during the knowledge acquisition process. R E F E R E N C E S F I. Angluin, D. (1988). Queries and concept learning. Machine Learning, 2, 319-342. 2. Bainbridge, L. (1986). Asking questions and accessing knowledge. In Future computing systems. New York: Elsevier. 3. Boose, J. (1984). Personal construct theory and the transfer of human expertise. Proc. of the Nat'l Conf on Artificial Intelligence, Austin, TX. 4. Boose, J. & Gaines, B. R. (1989). Knowledge acquisition for knowledge-based system: Notes on the state-of-the-art. Machine Learning, 4, 131-143. 5. Braun, H. & Chandler, J. S. (1987). Predicting stock market behavior through rule induction: an application of the learning- from-example approach. Decision Sciences, 18, 415-429. 6. Buchanan, B. G. et al. (1983). Constructing an expert system. In E Hayes-Roth, D. Waterman & D. Lenat (Eds), Building expert systems. Reading, Addison-Wesley. 7. Buntine, W. & Stirling, D. (1990). Interactive induction. In J. E. Hayes-Michie, D. Michie & E. Tyugu (Eds), Machine intelligence, Vol. 12, pp. 121-138, Oxford: Oxford University Press. 8. Carter, C. & Catlett, J. (1987). Assessing credit card applications using machine learning. IEEE Expert, 2, 71-79. 9. Cohn, D., Atlas, L. & Ladner, R. (1994). Improving generalization with active learning. Machine Learning, 15, 201-221. 10. Chung, H. M. & Silver, M. S. (1992). Rule-based expert systems and linear models: A n empirical comparison of learning-by- example methods. Decision Science, 23, 687-707. 11. Eriksson, H. (1992). A survey of knowledge acquisition techniques and tools and their relationship to software engineering. Journal of System and Software, 19, 97-107. 12. Evans, B. & Fisher, D. (1994). Overcoming process delays with decision tree induction. IEEE Expert, 60--66. 13. Feigenbaum, E. A. (1977). The art of artificial intelligence: Themes and case studies of knowledge engineering. International Joint Conference on Artificial Intelligence, Vol. 5, pp. 1014-1029. 14. Gaines, B. R. (1988). Knowledge acquisition: Development and advances.. In M. D. Oliff, (Ed.) Expert system and intelligent programming. New York: Elsevier. 15. Goldman, S. A. & Solan, R. H. (1994). The power o f self-directed learning. Machine Learning, 14, 271-294. 16. Hart, A. (1992). Knowledge acquisition for expert systems, 2nd edn. New York: McGraw-Hill. 17. Hayes-Roth, E , Waterman, D. A. & Lenat, D. (1983). Building expert systems. Reading MA: Addison-Wesley. 18. Jeng, B. & Weyuker, E. (1994). A simplified approach to domain testing. ACM Transactions on Software Engineering and Method- ology, 3, 254-270. 19. Jeng, B., Liang, T. P, & Jeng, Y. M. FILM: A fuzzy inductive learning method for automatic knowledge acquisition. Forthcoming in Decision Support Systems. 20. Johnson, P. E. (1983). W h a t kind of expert should a system be? Journal of Medicine and Philosophy, g, 77-97. 21. Kelly, G. A. (1955). The psychology of personal constructs., New York: Norton. 22. Liang, 1". P. (1992). A composite approach to inducing knowledge for expert systems design. Management Science, 3g, 1-17. Interactive Induction o f Expert Knowledge 401 23. Marchand, A., VanLente, E & Galen, R. (1983). The assessment of laboratory tests in the diagnosis of acute appendicitis. American Journal of Clinical Pathology, 80:3, 369-374. 24. Marcus, S. (1988). Automating knowledge acquisition for expert systems. Boston, MA: Kluwer. 25. McGraw, K. L. & Harbison-Briggs, K. (1989). Knowledge acquisition: Principles and guidelines. Englewood Cliffs, NJ: Prentice-Hall. 26. Messier, W. E & Hansen, J. V. (1988). Inducing rules for expert system development: an example using default and bankruptcy data. Management Science, 34, 1403-1415. 27. Michalski, R. S. & Chilausky, R. L. (1980). Knowledge acquisition by encoding expert rules versus computer induction from examples: A case study involving soybean pathology. Int. J. Man-Machine Study, 12, 63-87. 28. Michie, D. (Ed.) (1984). Introductory readings in expert systems. New York: Breach. 29. Michalski, R., Mozetic, I., Hong, J. & Lavrac, N. (1986). The multi-purpose incremental learning system AQ25 and its testing application to three medical domains. In Proc. of the 5th Annual Nat'l Conference on Artificial Intelligence, pp. 2041-2045, Phil- adelphia, PA. 30. Mitchell, T. M. 0982). Generalization as search. Artificial Intelligence, lg, 203-226. 31. Parsaye, K. (1988). Acquiring and verifying knowledge automat- ically. A1 Expert, pp. 48-63. 32. Paterson, A. & Niblett, T. 0982). ACLS user manual. Scotland: Intelligence Terminal Ltd. 33. Porter, B.W., Bareiss, R. & Holm, R.C. (1990). Concept learning and heuristic classification in weak-theory domains. Artificial Intelligence, 45, 229-263. 34. Quinlan, J. R. (2979). Discovering rules from large collections of examples: a case study. In D. Michie (Ed.), Expert systems in the micro electronic age. Scotland: Edinburgh University Press. 35. Quinlan, J. R. (1986). Induction of decision trees. Machine Learning, 1, 81-106. 36. Rendell, L. A. (1986). A general framework for induction and a study of selective induction. Machine Learning, 1, 177-220. 37. Ruff, R. A. & Dietterich, T. G. (1989). What good are experiments. Proc. of the Sixth International Workshop on Machine Learning, pp. 109-222, Ithaca, New York. 38. Scott, P. D. & Markovitch, S. (1993). t~xperience selection and problem choice in an exploratory learning system. Machine Learning, 12, 49-67. 39. Shaw, M. J. & Gentry, J. A. (1988). Using an expert system with inductive learning to evaluate business loans. Financial Manage- ment, 17, 45-56. 40. Turban, E. (1992). Expert systems and applied artificial intelli- gence. New York: Macmillan. 41. Utgoff, P. E. (1989). Incremental induction of decision trees. Machine Learning, 4, 161-186. 42. Wardle, A. & Wardle, L. (1978). Computer aided diagnosis--a review of research. Meth. Inform. Med., 17, 15-28. 43. Winston, P. H. (1975). Learning structural descriptions from examples. In P. H. Winston (Ed.), The psychology o f computer vision. New York: McGraw-Hill.