Buchanan, B. G., & Mitchell, T. M. (1978). Model-directed learning of production rules. In Waterman & Hayes-Roth (Eds.), Pattern-directed inference systems. New York: Academic Press.  Home/Search   Document Not in Database   Summary   Related Articles   Check

.... the appropriate class (or concept) Learning problems of this type include: learning a checkers evaluation function [29, 30] that assigns to a given class of board situations a certain value; learning descriptions of block structures [36] determining rules for interpreting mass spectrograms [5]; formulating diagnostic rules for soybean diseases [25] and discovering heuristics to guide the application of symbolic integration operators [27] In Samuel s checkers program, for example, each training instance was a board situation represented as a vector of 16 attributes. The learned ....

....of line endings, closed contours, and so on (e.g. 14] These position invariant properties can be made explicit by applying task oriented transformations to the raw data. An example of a learning program that performs task oriented trans formations is INTSUM (a part of the Meta DENDRAL system [5]) INTSUM is presented with raw training instances in the form of chemical structures (graphs) and associated mass spectra (represented as fragment masses and their intensities) For each fragment in the mass spectrum, INTSUM must determine the bonds that could have broken to produce that ....

[Article contains additional citation context not shown here]

Buchanan, B.G. and Mitchell, T.M., Model-directed learning of production rules, in: D.A. Waterman and F..Hayes-Roth (Eds.), Pattern-directed In/erence Systems (Academic Press, New York, 1978).

....particular, M82] presents a unifying approach to the problem of generalizing knowledge (that, for example, can be represented as rules) by viewing generalization as a search problem. Moreover, based on the specific search methods used, Mitchell [M82] also categorizes several rule learning systems [BM78, P70, W75, HRM7, V78, M77, MUB82] that deal with generalization. In particular, M82] deals with a broader set of objects (that can also include rules) and formulates the generalization problem as follows. Given a set of instances specified in an instance language, the generalization problem is formulated in [M82] as a search for ....

Buchanan, B.G. and Mitchell, T.M., 1978. Model Directed Learning of Production Rules. In Waterman and Hayes-Roth (eds.) Pattern-Directed Inference Systems, Academic Press, New York.

....set of potential feature vectors. In our lock smith problem, instead of knowing which key (from each key chain) opens the supply room, the learning system only knows that one of the keys on the key chain opens the door. An early example of this learning situation arose in the Meta DENDRAL project (Buchanan Mitchell, 1978; Lindsay, Buchanan, Feigenbaum Lederberg, 1980) In Meta DENDRAL, the goal was to learn rules that could predict the behavior of molecules inside a mass spectrometer. A mass spectrometer bombards a molecule with high energy particles, which causes the molecule to break into fragments. These ....

Buchanan, B. G. & Mitchell, T. M. (1978). Model-directed learning of production rules. In D. A. Waterman & F. Hayes-Roth (Eds.), Pattern-Directed Inference Systems. New York: Academic Press. 297--312.

....approach taken by that system allows the efficient, automatic generation and evaluation of a large number of learning structures (i.e. decision trees) that can be naturally decomposed into production rules. Several other researchers have investigated instance to class learning problems (e.g. [BM78] [MC80] MMHL86] For a summary of current research directions, see [MAL 86] Lan87b] 10 4 The Software Environment Sixteen moderate and large size software systems were selected from a NASA software production environment for this study [BZM 77] CMP 82] SEL82] The software is ground ....

B. G. Buchanan and T. M. Mitchell. Model-directed learning of production rules. In D. A. Waterman and F. Hayes-Roth, editors, Pattern-Directed Inference Systems, Academic Press, New York, 1978.

.... Research toward this goal has produced a variety of learning algorithms (e.g. the Perceptron [Minsky Papert, 1972] ID3 [Quinlan, 1986] backpropagation [Rumelhart McClelland, 1986a] and some impressive performance programs (e.g. Samuel s checker program [Samuel, 1959] MetaDendral [Buchanan Mitchell, 1978], LEX [Mitchell, Utgoff Banerji, 1983] However, in spite of these successes, few of today s problem solving systems have the ability to learn. One reason that inductive learning algorithms are not more widely used in problemsolving systems is their sensitivity to the way in which information ....

Buchanan, B. G., & Mitchell, T. M. (1978). Model-directed learning of production rules. In Waterman & Hayes-Roth (Eds.), Pattern-directed inference systems. New York: Academic Press.

....experience. Research toward this goal has produced a variety of inductive learning algorithms (e.g. the Perceptron [Minsky Papert, 1972] AQ11 [Michalski Chilausky, 1980] ID3 [Quinlan, 1983] and some impressive performance programs (e.g. Samuel s checker program [Samuel, 1959] MetaDendral [Buchanan Mitchell, 1978], LEX [Mitchell, Utgoff Banerji, 1983] However, in spite of these successes, few of today s problem solving systems have the ability to learn. One possible reason that inductive learning algorithms are not more widely used in problem solving systems is their sensitivity to the way in which ....

Buchanan, B. G., & Mitchell, T. M. (1978). Model-directed learning of production rules. In Waterman & Hayes-Roth (Eds.), Pattern-directed inference systems. New York: Academic Press.

....and learning. Included in these early programs were some for mathematics (e.g. Newell, Simon and Shaw s LOGIC THEORIST [Newell Simon 1956] and Herbert Gelernter s THEOREM PROVER [McCorduck 1979, p. 106] Gelernter 1963] and, for science (e.g. Stanford s DENDRAL and meta DENDRAL projects [Buchanan Mitchell 1978][Lindsay et al. 1980] Some early programs had some noted successes. In mathematics, Lenat s AM rediscovered the fundamental theorem of arithmetic (the unique factorization into primes) and Goldbach s conjecture (every composite number greater than 2 is the sum of two primes) Lenat 1976] Lenat ....

Buchanan, B.G., Mitchell, T.M. [1978] Model-directed learning of production rules. In D.A. Waterman and F. Hayes-Roth (Eds.) Pattern-directed inference systems. New York: Academic Press, p. 297-312.

....behavior for each account, then they are combined in a rule selection step. 4.1.1. Rule generation DC 1 uses the RL program (Clearwater and Provost 1990; Provost and Aronis 1996) to generate indicators of fraud in the form of classification rules. Similar to other MetaDENDRAL style rule learners (Buchanan and Mitchell 1978; Segal and Etzioni 1994; Webb 1995) RL performs a generalto specific search of the space of conjunctive rules. This type of rule space search is described in detail by Webb (Webb 1995) In DC 1, RL uses a beam search for rules with certainty factors above a user defined threshold. The certainty ....

Buchanan, B. G. and T. M. Mitchell (1978). Model-directed learning of production rules. In F. Hayes-Roth (Ed.), Pattern-directed inference systems, pp. 297--312. New York: Academic Press.

....methods, an a priori model is used to constrain the search. These methods search a set of possible generalisations in an attempt to find a few best hypotheses that satisfy certain requirements. Typical examples of systems which adopt this strategy are AM [Len77] DENDRAL and Meta DENDRAL [BuM78], and the approach used in the INDUCE system [DiM81] Data driven techniques generally have the advantage of supporting incremental learning. The learning process can start not only from the specific training examples, but also from the rules which have already been discovered. The learning ....

B.G. Buchanan and T. M. Mitchell, (1978). Model-Directed Learning of Production Rules, Pattern-Directed Inference System, Academic Press, Waterman et. al. (eds), 291-312.

....from domain examples. Mitchell s version spaces [Mit77] algorithm learned by looking at examples one after the other and generalizing or specializing as necessary so as to arrive at a reasonable rule set. This system relied on a sequential analysis of carefully chosen examples. Meta DENDRAL [BM78] is an extension of the DENDRAL expert system which was able to learn its own production rules in its specific domain. It was quite successful, even discovering rules previously unknown to chemists, but did not generalize well to other domains. Michalski s AQ11 [MC80] was a more general ....

B. Buchanan and T. Mitchell, "Model-directed learning of production rules," In D. Waterman and F. Hayes-Roth, editors, Pattern-Directed Inference Systems, pages 297--312. Academic Press, New York, NY, 1978.

.... the disjunction WaitEstimate(x, 30 60) WaitEstimate(x, 60) we might use the single literal LongWait(x) The pure version space algorithm was first applied in the MetaDENDRAL system, which was designed to learn rules for predicting how molecules would break into pieces in a mass spectrometer (Buchanan Mitchell, 1978). MetaDENDRAL was able to generate rules that were sufficiently novel to warrant publication in a journal of analytical chemistry the first real scientific knowledge generated by a computer program. 3 Inductive logic programming Inductive logic programming (ILP) is one of the newest subfields ....

Buchanan, B. G., & Mitchell, T. M. (1978). Model-directed learning of production rules. In D. A. Waterman & F. Hayes-Roth (Eds.), Pattern-directed inference systems. New York: Academic Press.

....knowledge to re express data in terms of the higherlevel features used by molecular biologists. It is the definitions of these higher level features, extracted directly from the biology literature, that form the background knowledge that we use. Our efforts are thus similar to that of Meta DENDRAL [4], WYL [7] and MIRO [5] using domain specific knowledge to generate appropriate training data representations. We differ, however, in the development of domain specific training data representations that are independent of any specific learning method. This paper is structured as follows. Section ....

B. G. Buchanan and T. M. Mitchell. Model-directed learning of production rules. In D. A. Waterman and F. Hayes-Roth, editors, Pattern-Directed Inference Systems, pages 297--312. Academic Press, New York, 1978.

....representation for a problem automatically. Constructive induction has been investigated as a domain independent or as a domaindependent task. The former gave rise to systems such as stagger [ Schlimmer, 1987 ] fringe [ Pagallo, 1989 ] and citre [ Matheus Rendell, 1989 ] while metadendral [ Buchanan Mitchell, 1978 ] stabb [ Utgoff, 1986 ] duce [ Muggleton, 1987 ] miro [ Drastal, Czako, Raatz, 1989 ] ioe [ Flann Dietterich, 1990 ] cindi [ Callan Utgoff, 1991 ] and zenith [ Fawcett Utgoff, 1992 ] are illustrations of the latter. 1 Domain independent automated approaches are appealing in ....

Buchanan, B. G., and Mitchell, T. M. 1978. Model-directed learning of production rules. In Waterman, D. A., and HayesRoth, F., eds., Pattern-Directed Inference Systems. New York: Academic Press. 297--312.

....on the resulting APR. It estimates the density of instances along each feature of the APR, and expands the bounds of the APR so that the estimated probability of excluding a positive instance is #. Although Multiple Instance problems have been encountered before (for example, in Meta Dendral [ Buchanan and Mitchell, 1978 ] they were transformed into a traditional supervised learning problem. Dietterich et al. s paper is important because it was the first to give an algorithm specifically for learning from Multiple Instance examples, and also because it achieved impressive results on the musk data set. However, ....

....and not by negative ones. This is an ambiguous learning problem because a labeled example (DFA) represents a variety of di#erent strings. Unfortunately, Cohen shows that it is hard to PAC learn from these examples, even if the alphabet size is limited to three characters. The Meta DENDRAL program [ Buchanan and Mitchell, 1978 ] receives training pairs of molecules and their mass abundance curve. Its goal is to learn how molecules disintegrate during mass spectrometry which bonds are broken and which atoms migrate. Since every location along the mass abundance curve can be explained by many di#erent hypotheses, the ....

B. G. Buchanan and T. M. Mitchell. Model-directed learning of production rules. In D. A. Waterman and F. Hayes-Roth, editors, Pattern-Directed Inference Systems. Academic Press, 1978.

....The fascination with this ability has attracted many researchers in an attempt to exploit learning when building artifacts such as computer programs. Initially, these efforts concentrated in specific large expert systems projects (e.g. Teiresias for MYCIN (Davis, 1979) RULEGEN for Meta Dendral (Buchanan and Mitchell, 1978), and later, LEAP for VEXED (Mitchell et al., 1985) and their objective was very practical to solve the knowledge acquisition bottleneck of building these systems. Over time and in a bottom up fashion, two major research thrusts have emerged: 1) creating computer programs for solving some ....

Buchanan, B. G. and Mitchell, T. M. (1978). "Model-directed learning of production rules." In Waterman, D. A. and Hayes-Roth, F., editors, Pattern-Directed Inference Systems, New York, NY, Academic Press.

....mining, involves combing through the call data searching for indicators of fraud. In the DC 1 system, the indicators are conjunctive rules discovered by a standard rule learning program. We use the RL program (Clearwater Provost 1990) which is similar to other MetaDENDRAL style rule learners (Buchanan Mitchell 1978; Segal Etzioni 1994) RL searches for rules with certainty factors above a user defined threshold. The certainty factor we used for these runs was a simple frequency based probability estimate, corrected for small samples (Quinlan 1987) The call data are organized by account, and each call ....

Buchanan, B. G., and Mitchell, T. M. 1978. Modeldirected learning of production rules. In Hayes-Roth, F., ed., Pattern-directed inference systems. New York: Academic Press.

....of either the problem solver s decisions (causal dynamics) or the passage of time (natural dynamics) or both. Difficult projection problems result in the last case. Learning systems developed by artificial intelligence (AI) researchers have limited success in coping with natural dynamics [6, 7, 28, 31, 37, 43, 53, 99, 130, 136, 142, 147, 150, 152]. When learning strategies by trial and error, a learning system may experiment with multiple strategies (candidates) The measured performance of tested candidates provides feedback for guiding both the selection of candidates for future tests and the modification of incumbents toward improved ....

B. G. Buchanan and T. M. Mitchell, "Model-Directed Learning of Production Rules," in PatternDirected Inference Systems, ed. D. A. Waterman and F. Hayes-Roth, Academic Press, New York, NY, 1978.

....rather than taken from an actual trace of HAMB, but it is closely patterned after actual traces generated by HAMB. 3. The HAMB program As described in [1] HAMB uses rule induction as its primary method for discovering patterns and uses the rule induction program RL [3] derived from MetaDENDRAL [4]. Because HAMB mainly uses rule induction to perform discovery, HAMB s items are attributes, examples, rule conjuncts, and rules, plus sets of these. HAMB s discoveries are items with interesting properties or relationships. HAMB estimates the interestingness of its items and discoveries using a ....

Buchanan, B. G. and Mitchell, T. 1978. ModelDirected Learning of Production Rules. In Waterman, D. A. and Hayes-Roth, F., Eds., Pattern Directed Inference Systems, 297-312. New York, NY: Academic Press.

....the first attribute to partition the cases to the entropy of the first attribute s values (c.f. Quinlan s information gain [5] 3. The HAMB program HAMB uses rule induction as its primary method for discovering patterns and uses the rule induction program RL [6] a descendant of Meta DENDRAL [7]. RL is efficient, flexible, robust with respect to incomplete or noisy data, and uses a variety of domain knowledge. Because HAMB uses rule induction to perform empirical discovery, HAMB s items are attributes, examples, rule conjuncts, and rules, plus sets of any of these. HAMB s discoveries ....

Buchanan, B. G. and Mitchell, T. 1978. Model-Directed Learning of Production Rules. In Waterman, D. A. and Hayes-Roth, F., Eds., Pattern Directed Inference Systems, 297-312. New York, NY: Academic Press.

....the proposed agenda and justification based discovery framework we implemented it in a prototype discovery program called HAMB. HAMB uses rule induction as its primary method for discovering patterns and uses the rule induction program RL [Provost and Buchanan, 1995] a descendant of Meta DENDRAL [Buchanan and Mitchell, 1978]. RL is efficient, flexible, robust with respect to incomplete or noisy data, and can use a variety of domain knowledge [Clearwater and Provost, 1990; Provost and Buchanan, 1995] Figure 1 sketches HAMB s top level control and Figure 2 presents an overview of the types of tasks HAMB performs. ....

Bruce G. Buchanan and Tom Mitchell. Model-directed Learning of Production Rules. In D. A. Waterman and F. Hayes-Roth, Eds.,

.... of early uses of search for rule induction, culminating in the de nition of version spaces (described below) Probably the rst successful application of rule space search for knowledge discovery was in the Meta DENDRAL program (Buchanan, Feigenbaum, Lederberg, 1971; Buchanan Feigenbaum, 1978; Buchanan Mitchell, 1978), which performed what would now be called data mining for scienti c discovery in Chemistry. Meta DENDRAL used chemistryspeci c knowledge for pruning, and not only rediscovered known, published rules of mass spectrometry, but also made novel discoveries that were published in the chemistry ....

....set of most speci c rules in the version 10 space (the S set) together are sucient to describe the entire version space. Hirsh relaxes Mitchell s requirement of the rules strict consistency with the training data. In fact, the notion of version spaces arose from early work on rule space search (Buchanan Mitchell, 1978). Because the rules in the entire version space are logically entailed by the G set, and any given rule in the G set can subsume many di erent satisfactory specializations in the version space, we have found that in many domains the G set contains the most interesting rules. For instance, if the ....

Buchanan, B., & Mitchell, T. (1978). Model-directed learning of production rules. In Waterman, D., & Hayes-Roth, F. (Eds.), Pattern Directed Inference Systems. Academic Press., New York, NY.

....or stronger reasons has a greater priority . A task proposed more than once with different reasons has a greater priority . A task involving a concept with a higher interestingness has a greater priority RL and rule induction. HAMB uses the rule induction program RL, a descendant of MetaDENDRAL (Buchanan and Mitchell 1978), as its primary generator of new concepts. RL s output is a disjunctive set of conjunctive rules, each of which has the form: IF R 1 (A 1 ,V 1 ) R n (A n ,V n ) THEN Class Membership(Class k ) yes no, where each conjunct in its left hand side (LHS) names a simple relationship between the ....

Buchanan, B. G. and Mitchell, T. 1978. Model-directed Learning of Production Rules. In Waterman, D. A. and Hayes-Roth, F., Eds., Pattern Directed Inference Systems. New York, New York: Academic Press, 297-312.

No context found.

Buchanan, B. G., & Mitchell, T. M. (1978). Model-directed learning of production rules. In Waterman & Hayes-Roth (Eds.), Pattern-directed inference systems. New York: Academic Press.

No context found.

B. Buchanan and T. Mitchell. Model-directed learning of production rules. In Waterman, D. A. and Hayes-Roth, F. (Eds.) PatternDirected Inference Systems, Academic Press, New York, 1978.

No context found.

B.G.Buchanan & T.M.Mitchell; Model-Directed Learning of Production Rules. In Pattern Directed Inference Systems, Waterman & Hayes-Roth (Eds), 1978, 297-312

First 50 documents

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC