Quinlan, R. (1990) Learning Logical Definitions from Relations, in: Machine Learning, 5, 239-266.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....classification theories. In the framework of Inductive Logic Programming (ILP) a number of methods have been devised to solve the problem: transformation of relational problems into equivalent propositional ones as in LINUS [18] use of a priori knowledge either in a procedural form as in FOIL [25] or in a declarative form as in Progol [22] The first class of methods includes the work by Dzeroski et al. 12] who propose the transformation of first order representations into propositional form, in order to handle real numbers by means of techniques already tested in decision tree ....

...., thus we have to complete it by adding a further clause that covers at least e 2 . In the separate stage both learning systems search in the space of all value definitions for a complete model. The separate and conquer search strategy is adopted in other well known learning systems, such as FOIL [25]. The main difference with our systems is the use of a seed example, whose function is that of guiding the learning process. Since each positive example should be covered by at least one clause of the model returned by the procedure separate and conquer, each example becomes a valuable source of ....

R. Quinlan, Learning logical definitions from relations, Machine Learning 5 (1990), 239--266.

.... outputs : C : clause) Select a predicate p that must be learned Initialize C to be p(X) repeat (Specialization loop) Find the refinement C best 2 ae(C) according to some heuristic function Assign C : C best until Necessity stopping criterion is satisfied return C FOIL [Qui90b] mFOIL [Dze91] and Progol [Mug95a] are examples of systems based on this algorithm. The algorithm starts with an empty hypothesis H and a current set of example E cur that is initially set to the entire training set. The algorithm is composed of two repeat loops, referred to as covering and ....

....as relative least general generalization (rlgg) Plo70] see section 3.2.5) and the GOLEM system [MF90] see section 3.4.1) inverse resolution [MB92] or inverse entailment [LM92] Conversely, MGSs can be found by adopting a top down refining method (see section 3.2. 6) and a system such as FOIL [Qui90b] see section 3.4.2) or Progol [Mug95a] 88 5.4 Strategies for Combining Different Generalizations The generality of concepts to be learned is an important issue when learning in a threevalued setting. In a two valued setting, once the generality of the definition is chosen, the extension ....

[Article contains additional citation context not shown here]

J.R. Quinlan. Learning logical definitions from relations. Machine Learning, 5:239--266, 1990.

....to the absence of some facts related to them in the background. This may require the learning of multiple overspecific rules for covering a set of examples that could otherwise be covered by a single general one. Various systems have been developed to learn from imperfect data (for example, FOIL [Qui90a] mFOIL [Dze91] FOIL I [IKI 96] and LINUS [LDG91b] However, no system has been specially designed for learning from an incomplete background knowledge.This problem can be solved by integrating abductive reasoning into induction: abduction is used in order to complete the background ....

....can thus be used to classify new unseen examples that are incompletely specified. The system has been tested on a number of datasets where the knowledge is incomplete and the results obtained have been compared with those of state of the art systems like ICL Sat [DRD96c] mFOIL [Dze91] and FOIL [Qui90a] The performances of ACL were found to be superior or comparable with those of these systems on the considered datasets. By means of Extended logic programs we are able to represent and reason with information in a three valued logical setting. Extended logic programs contain two kinds of ....

[Article contains additional citation context not shown here]

J. R. Quinlan. Learning logical definitions from relations. Machine Learning, 5:239--266, 1990.

....cover single examples. Regarding compression measures for hypotheses, those clauses are discarded. Still, they carry valuable information: One could imagine, that we collect all literals from all redundant clauses and then construct a new clause using an information gain driven method (see Foil, [10]) ....

Quinlan, J. R. Learning logical definitions from relations. Machine Learning 5, 3 (1990), 239 -- 266.

....wide variety of learning models and methods. Some of these are associated with the field of connectionism, e.g. backpropagation [1] others are associated with work on evolutionary and genetic methods [2] still others are associated with work in the field of machine learning, e.g. C4.5 [3] Foil [4] and Golem [5] Given the wide range of methods on offer, it is important for researchers to adopt a principled approach to the task of matching up methods with learning problems. Unfortunately, there are still major obstacles to be overcome. A central question in any attempt to find a solution ....

Quinlan, J. (1990). Learning logical definitions from relations. Machine Learning, 5 (pp. 239-266).

....(for example, learning recursive concepts) but is also generally believed to be very inefficient for large databases. Blockeel et al. 1999] developed a scalable ILP system named TILDE that effectively flattens relational data into what they call interpretations and then uses a version of FOIL [Quinlan, 1990] , modified to make efficient access to data from disk, on these interpretations. TILDE was evaluated on a synthetic data set with 100,000 training examples. VFREL scales to much larger data sets by using sampling to focus on relevant attributes and relations. Slattery and Craven [2001] ....

J. R. Quinlan. Learning logical definitions from relations. Machine Learning, 5:239--266, 1990.

....In this section we will have a look at two di#erent approaches for inducing Horn clauses. The first approach is FOIL, a top down algorithm. The second one is relative least general generalization, which is a bottom up approach. 4.3. 1 FOIL The first algorithm we will take a look at is FOIL [Qui90]. The pseudo code for the algorithm can be found in figure 4.3. FOIL is a quite naive algorithm that starts with a general rule and will add literals to that rule until it no longer covers any of the negative examples. It will create a (sometimes extremely large) collection of literals which it ....

J.R. Quinlan. Learning logical definitions from relations. Machine Learning, 5:239--266, 1990.

....growth of Internet requires tools and methodologies developed to manage this increasing and changeable bulk of information, one of them is information filtering. It is recent specialization within information science dedicated to searching, consulting and retrieving of useful information [ 1, 2, 17]. The problem of Web search is translated into the problem of data mining [3, 4] without making an exhaustive revision that would take a user a lot of time, which is in many cases restricted. A number of approaches to the discovering of pages of interest have been developed based on the ....

....articles, etc. and ordinal variables that correspond to the classification categories of the pages [5, 7] In these approaches, the arithmetic handling of the frequencies of key occurrences of words is used, which is a very empiric way of classification, though computationally very efficient [2, 7, 9]. These approaches are close to the crisp Bayesian approach concerning the probabilities of accepting some page a posteriori, based on the existing evidence. However, the fact that one page is interesting or important for the user corresponds to an event of fuzzy nature, this is, the fact that ....

[Article contains additional citation context not shown here]

Quinlan,J., Learning logical definitions from relations. Machine Learning 5:239-266. 1993.

....we introduce the concept of a clause description CD(e t ) for an example representation e t . Let be the set of predicates introduced in Section 4.1. We call this the hypothesis language which is used later for construction of rules. This is in analogy to standard ILP algorithms like FOIL [Quinlan, 1990] . It should be noted that the hypothesis language can be freely chosen. Furthermore let us assume that a logic program is given that implements the intended semantics of the predicates in . To denote the union of and (D i ) we write P . Now we can define CD(e t ) l l # ....

J. R. Quinlan. Learning logical definitions from relations. Machine Learning, 5:239--266, 1990.

....algorithm builds the basis for many inductive rule learning algorithms it is in combination with CN2 a propositional rule learner. Nevertheless it builds a basis for many first order rule learning approaches. A widely used first order rule learning algorithm is the top down procedure called FOIL [Quinlan, 1990] . Modified versions of FOIL are the basis for most adaptive IE systems using relational learning techniques. For example, the SRV system [Freitag, 1998] uses a FOIL based core algorithm. FOIL tries to find a description of a target predicate, given a set of examples and some background knowledge. ....

J. R. Quinlan. Learning logical definitions from relations. Machine Learning, 5:239--266, 1990.

....The extension is determined concretely from the intensional description and an information ordering is defined over the feature structures. Returning to machine learning, algorithms such as decision trees [5, 6] the version space [7, 8] algorithm and techniques of inductive logic programming [9], formulate intensional descriptions of classes of objects given a number of training examples. These algorithms are intensional because they characterise classes in terms of attribute and predicate expression based descriptions intended to be comprehended by a human being. In summary, the ....

....storing the complement of extension of a term is trivially converted into a sequence of intervals containing document numbers of documents possessing the term. For example, if the term m is not present in documents 1, 7, 8, 13, 15, 19 then we can generate the sequence of intervals [1 6] [9 12], 14 14] 16 18] The algorithm split across Figures 2.18 and 2.19 follows the same principle as the algorithm in Figure 2.16 but makes use of intervals. The heap, T , now stores interval attribute pairs. The ordering on the pairs is still generated from the document interval part of ....

J. Quinlan, "Learning logical definitions from relations," tech. rep., University of Sydney, 1989.

.... classification machine learning algorithms, including decision trees, such as C4.5[72] CART (Classification and Regression Tree [14] and CHAID (Chi Square decision tree) 28] neural networks [31] genetic algorithms [28] bayesian based methods [57] classification rule algorithms, such as FOIL [71], Ripper, and more. In addition to these base algorithms, several approaches for developing committees, or combinations of learners have been used to improve learning [57, 31] Each of these algorithms has its advantages and disadvantages, depending on a given problem. See [57] for an overview of ....

Quinlan, J.R., "Learning logical definitions from relations", Machine Learning , vol. 5, pp. 239-266.

....given R . The system was generally able to complete correctly all 112 triplets even when 28 of them, picked at random, had been left out during training. These results on the Family Tree Problem are much better than the ones obtained using any other method on the same problem: Quinlan s FOIL [7] could generalize on 4 triplets, while Hinton (1986) and O Reilly (1996) made one or more errors when only 4 test cases were held out during training. The names of the Italian family have been altered from those originally used in Hinton (1986) to match those of one of the author s family. For ....

.... we have shown elsewhere [4] that, after learning a distributed representation for a set of concepts and relations, LRE can easily modify these representations to incorporate new concepts and relations and that it is possible extract logical rules from the solution and to couple LRE with FOIL [7]. Learning is fast and LRE rarely converges to solutions with poor generalization. We began introducing LRE for binary relations, and then we saw how these ideas can be easily extended to higher arity relation by simply concatenating concept vectors and using rectangular matrices for the ....

J. R. Quinlan. Learning logical definitions from relations. Machine Learning, 5:239--266, 1990.

....or implication [13] The objective of this work is to extend the result on the operators for clauses and prove the existence of ideal refinement operators for spaces of theories [9, 1] in those generalization models. Indeed, most algorithms for relational learning, such as those employed in FOIL [15] and PROGOL [11] adopt greedy iterative covering strategies based of the refinement of clauses. Although these refinements may turn out to be optimal for a single clause, the result of assembling them in a theory is not guaranteed to be globally effective, since the interdependence of the clauses ....

J. Quinlan. Learning logical definitions from relations. Machine Learning, 5:239--266, 1990.

.... or semantic fragments in order to guess design constraints and assumption about the database ( And94] PKBT94] A survey of those methods can be found in [Hai91] 10.3 Machine Learning Techniques Decision trees are introduced in [Quia] Inductive Logic Programming is introduced with FOIL in [Quib]. FOIL is a top down system, it starts with the most general rules and learns by specializing. Bottom up logic programming is introduced with GOLEM in [MF92] Learning in GOLEM proceeds by generalizing. Version Spaces as a learning algorithm is introduced in [Mit82] Version Spaces are studied ....

J.R. Quinlan. Learning logical definitions from relations. Machine Learning.

....and they also facilitate debugging during software development. The issue of integrating learning into inference systems has been studied intensely before. For example, recent work on explanation based learning [69, 42, 20] theory refinement [94, 108, 80, 78] and inductive logic programming [73, 85] has led to a variety of learning algorithms that modify programs written in first order logic based on examples. Several research teams have integrated such learning algorithms into problem solving architectures, such as SOAR [89, 26, 66, 57] PRODIGY [67, 39] and THEO [70] These architectures ....

J. R. Quinlan. Learning logical definitions from relations. Machine Learning, 5:239--266, 1990.

....The SmL learner for mDTD is based on the previous CN2 algorithm [4] except for the measure of performance evaluation. Figure 1 shows the SmL algorithm which uses a general to specific beam search. In CN2, the performance measure for the generated rule is an information gain, similar to FOIL [7], while, in SmL, lexical similarity and coverage rate for examples are newly introduced as a measure in order to efficiently process the heterogeneous textual data in Web sites. The set of positive examples for learning is simply a set of field instances automatically extracted from structured ....

J. Quinlan, Learning Logical Definitions from Relations, Machine Learning, Vol. 5, 1990.

....for generality have been proposed within the field of inductive logic programming. These include for instance inverse implication, inverse resolution and inverse entailment (see [Muggleton and De Raedt, 1994] for an overview) However, in practice, the large majority of ILP systems (including FOIL [Quinlan, 1990], GOLEM [Muggleton and Feng, 1990] PROGOL TM [Muggleton, 1995] CLAUDIEN [De Raedt and Dehaspe, 1997a] and TILDE [Blockeel and De Raedt, 1998; Blockeel, 1998] uses 0 subsumption. This is due to the better computational properties of 0 subsumption as compared to inverse resolution and inverse ....

....in [Driessens et al. 2001] based on the propositional G tree algorithm in [Chapman and Kaelbling, 1991] 4. 10.2 Others There are plenty of other practical inductive logic programming and relational data mining systems which can be seen as an upgrade of a propositional learner, e.g. FOIL [Quinlan, 1990] can be considered an upgrade of AQ [Michalski t at. 1986] or CN2 [Clark and Niblett, 1989; Clark and Boswell, 1991] see also below) RIBL [Emde and Wettschereck, 1996] upgrades the classical k nearest neighbor algorithm (using a first order distance due to IBisson, 1992] SRT [Kramer, ....

[Article contains additional citation context not shown here]

J.R. Quinlan. Learning logical definitions from relations. Ma- chine Learning, 5:239-266, 1990.

....being appropriate as learning tools. Instead, what is required is inductive logic programming that learns not only attributes and propositional descriptions, but also predicates for finding useful relations in obscure combinations. In the experiment, the authors use a learner similar to FOIL [ Quinlan, 1990 ] except that its background knowledge may also be described using Horn clauses. 4.2 Background Knowledge for ILP The model is constructed based on background knowledge definitions of predicates that describe melody and chords. Melody is considered to be a sequence of notes that have pitch ....

J. R. Quinlan. Learning logical definitions from relations. Machine Learning, 5:239--266, 1990.

....prove examples. It has been applied to single predicate learning in the system ICN. 1 Introduction In the field of Inductive Logic Programming (I.L.P. many works deal with the task of learning a definition of a concept from positive and negative examples of this concept and a knowledge base [15, 16, 10, 6, 11]. The predicates specified in the knowledge base are called the basic predicates, and the predicate that must be learned is called the target predicate. The goal is to learn a concept definition, which is complete, i.e. which proves that all the positive examples are true , and consistent, i.e. ....

....predicate q. P proves that a ground atom e is true if P [ bk j= e, otherwise P proves that e is false. In the field of I.L.P. two main approaches can be distinguished: ffl interactive theory revision, illustrated by the systems MIS [20] and Clint [16] ffl empirical learning, illustrated by FOIL [15] and Golem [10] In the field of interactive theory revision, the systems check, each time a new clause is generated, whether the whole set of clauses remains complete and consistent w.r.t. the set of positive and negative examples that have already been processed. On the other hand, many ....

[Article contains additional citation context not shown here]

Quinlan J.R., 1990. Learning Logical Definitions from Relations. Machine Learning Journal, Vol. 5, Kluwer Academic Publishers, pp. 239-266.

No context found.

Quinlan, J.R., Learning logical definitions from relations. Machine Learning 5 (1990) 239-266.

No context found.

Quinlan, R. (1990) Learning Logical Definitions from Relations, in: Machine Learning, 5, 239-266.

No context found.

J.R. Quinlan. Learning logical definitions from relations. Mach. Learn., 5:239--266, 1990.

No context found.

J. R. QUINLAN (1990). Learning Logical Definitions from Relations. Machine Learning. 5(3), p. 239266.

No context found.

Quinlan, J.R. (1990): Learning logical definitions from relations. Machine Learning 5, 239-266.

No context found.

Quinlan, J. R. (1990). Learning logical definitions from relations. Machine Learning, 5:239--266. 231

No context found.

Quinlan, J. R. 1990. Learning logical definitions from relations. Machine Learning 5(3):239--266.

No context found.

Quinlan, J. R. (1990). Learning logical definitions from relations. Machine Learning, 5, 239--266.

No context found.

J.R. Quinlan. Learning logical definitions from relations. Machine Learning, 5:239-- 266, 1990.

No context found.

Quinlan, J. R. (1990). Learning Logical Definitions from Relations. Machine Learning, 5, 239--266.

No context found.

J.R. Quinlan. Learning logical definitions from relations. Machine Learning, 5(3):239--266, 1990.

No context found.

J. R. Quinlan. Learning logical definitions from relations. Machine Learning, 5:239--266, 1990.

No context found.

J. R. Quinlan. Learning logical definitions from relations. 5:239--266, 1990.

No context found.

Quinlan, J.R.; "Learning logical definitions from relations." Machine Learning, n. 5, pp.239-266.

No context found.

J. R. Quinlan. Learning logical definitions from relations. Machine Learning, 5:239--266, 1990.

No context found.

J.R. Quinlan, Learning logical definitions from relations, Machine Learning 5 (1990) 239-- 266.

No context found.

J. R. Quinlan. Learning logical definitions from relations. Machine Learning, 5(3):239--266, August 1990.

No context found.

J. Quinlan (1990). Learning logical definitions from relations. Machine Learning, 239266.

No context found.

Quinlan, J. R., "Learning Logical Definitions from Relations, " Machine Learning, 5(3), pp. 239-266, 1990.

No context found.

J. R. Quinlan. Learning logical definitions from relations. 5:239--266, 1990.

No context found.

Quinlan, J.R.: Learning Logical Definitions from Relations. Machine Learning 5 (1990) 239-266

No context found.

Quinlan, J. R. 1990. Learning logical definitions from relations. Machine Learning 5(3):239--266.

No context found.

J.R. Quinlan. Learning logical definitions from Relations. Machine Learning 5:239266, 1990.

No context found.

J.R. Quinlan. Learning logical definitions from relations. Machine Learning, 5:239-- 266, 1990.

No context found.

J. Quinlan (1990). Learning logical definitions from relations. Machine Learning, 5(3):239-266.

No context found.

J. R. Quinlan. Learning logical definitions from relations. 5:239--266, 1990.

No context found.

J. R. Quinlan. Learning logical definitions from relations. Machine Learning, 5(3):239--266, 1990.

No context found.

J. R. Quinlan. Learning logical definitions from relations. 5:239--266, 1990.

No context found.

R. Quinlan. Learning logical definitions from relations. Machine Learning, 5:239--266, 1990.

No context found.

Quinlan, J. (1990). Learning logical definitions from relations. Machine Learning, 5 (pp. 239-266). 6

First 50 documents  Next 50

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