J. M. Zelle and R. J. Mooney, J. B. Konvisser, Combining Top-down and Bottom-up Methods in Inductive Logic Programming, Proc of The 11th International Conference on Machine Learning (ML-94), pp.343--351, 1994.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....form of head :0 body. Covered examples are removed from P in each cycle. The inner loop consists of two phases: the head construction phase and the body construction phase. It constructs heads in a bottom up manner and constructs the body in a top down manner, following the result described in (Zelle et al. 1994). The search heuristic PWI is weighted informativity employing the Laplace estimate. Let T = fHead :0Body g [ BK. PW I(P; T ) 0 1 j P j 2 log 2 j P j 1 jQ(T )j 2 ; where j P j denotes the number of positive examples covered by T and Q(T ) is the empirical content. The smaller the ....

J. M. Zelle and R. J. Mooney, J. B. Konvisser, Combining Top-down and Bottom-up Methods in Inductive Logic Programming, Proc of The 11th International Conference on Machine Learning (ML-94), pp.343--351, 1994.

....all the learned models into a single model by Bayesian combination (Buntine, 1990) 3 Learning Multiple Models via Tabulate In this section, we are going to discuss the details of the new induction algorithm used by Chill. Since its design is strongly motivated by the Chillin induction algorithm (Zelle Mooney, 1994), we are going to provide a brief overview of the algorithm here. Then, we will proceed to explain the motivation for the new algorithm and discuss its details and various ILP issues it addresses. 3.1 Combining Top down and Bottom up Methods in Chillin Top down or bottom up approaches to ILP ....

....by the clause C is a subset of that of the clause 7 Chillin stands for the CHILL INduction algorithm. 16 D. 3. 2 Motivation for the New Algorithm Chillin has been tested on learning simple list processing programs (like append 3) and was shown to be more effective than either Foil or Golem (Zelle Mooney, 1994). While Chillin was shown to be a successful attempt to combining top down and bottom up approaches in ILP, it suffers from several weaknesses. First, the search is basically a hillclimbing search in the hypothesis space which may sometimes get stuck on a local minimum. It may be interesting to ....

Zelle, J. M., & Mooney, R. J. (1994). Combining top-down and bottom-up methods in inductive logic programming. In Proceedings of the Eleventh International Conference on Machine Learning, pp. 343--351 New Brunswick, NJ.

....greatest coverage of positive examples is selected, and that clause is further generalized by computing the rlggs of the selected clause with new randomly chosen positive examples. The generalization process stops when the coverage of the best clause no longer increases. 2.2. 3 Chillin The Chillin (Zelle Mooney, 1994) system combines top down (general to specific) and bottom up ILP techniques. The algorithm starts with a most specific definition (the set of positive examples) and introduces generalizations which make the definition more compact. Generalizations are created by selecting pairs of clauses in the ....

....constraint, and the constraint would be replaced by a semantic constraint. This semantic class would then be available to the system as it generalized other rules. The development of this new background knowledge is analogous to predicate invention in ILP (Kijsirikul, Numao, Shimura, 1992; Zelle Mooney, 1994). It would be interesting also to attempt to incorporate new words into an existing semantic class. For example, we might have a semantic class consisting of programming languages which we would like to expand automatically to incorporate new languages. Another addition we plan for the algorithm ....

Zelle, J. M., & Mooney, R. J. (1994). Combining top-down and bottom-up methods in inductive logic programming. In Proceedings of the Eleventh International Conference on Machine Learning, pp. 343--351 New Brunswick, NJ.

.... learn everything, and do without any background knowledge Statistical approaches such as SPATTER (Magerman, 1995) or the Bigram Lexical Dependency based parser (Collins, 1996) produce parsers on very limited linguistic background information and Inductive Logic Programming systems such as CHILL (Zelle Mooney, 1994; Zelle, 1995) have even generated linguistically relevant categories such as animate. In short, the background knowledge we use is at least qualitatively easy to provide and mostly already conceptually available in the form of traditional (paper) dictionaries and grammar books; using available ....

Zelle, J. M., & Mooney, R. J. (1994). Combining top-down and bottom-up methods in inductive logic programming. In Proceedings of the Eleventh International Conference on Machine Learning, pp. 343--351 New Brunswick, NJ.

....in a theory is re ned or 2) a new clause is begun. Clauses are learned using both top down specialization using a method similar to Foil (Quinlan, 1990) and bottom up generalization using Least General Generalizations (LGG s) Advantages of combining both ILP approaches were explored in Chillin (Zelle and Mooney, 1994), an ILP method which motivated the design of Tabulate. An outline of the Tabulate algorithm is given in Figure 2. A noise handling criterion is used to decide when an individual clause in a hypothesis is suciently accurate to be permanently retained. There are three possible outcomes in a re ....

J. M. Zelle and R. J. Mooney. 1994. Combining top-down and bottom-up methods in inductive logic programming. In Proceedings of the Eleventh International Conference on Machine Learning, pages 343-351, New Brunswick, NJ, July.

....p.7 8 predicates. In particular, several ILP methods for inventing predicates use abduction to infer training sets for an invented predicate and then invoke induction recursively on the abduced data to learn a definition for the new predicate (Wirth and O Rorke, 1991; Kijsirikul et al. 1992; Zelle and Mooney, 1994; Stahl, 1996; Flener, 1997) This technique is basically the same as using abduced data to learn new rules for existing predicates in theory refinement as described above. A final interesting point is that the same approach to using abduction to guide refinement can also be applied to ....

Zelle, J. M. and R. J. Mooney: 1994, `Combining Top-Down and Bottom-Up Methods in Inductive Logic Programming'. In: Proceedings of the Eleventh International Conference on Machine Learning. New Brunswick, NJ, pp. 343--351.

No context found.

Zelle, J. M., & Mooney, R. J. (1994a). Combining top-down and bottom-up methods in inductive logic programming. In Proceedings of the Eleventh International Conference on Machine Learning, pp.

.... such as prepositional phrase attachment and lexical ambiguity (Zelle and Mooney, 1993b) Fourth, it uses a single, uniform parsing framework to perform all of these tasks and a single, general learning method that has also been used to induce a range of diverse logic programs from examples (Zelle and Mooney, 1994). The remainder of the paper is organized as follows. In section 2, we summarize our ILP method for learning deterministic parsers, and how this method was tailored to work with existing treebanks. In section 3, we present and discuss experimental results on learning parsers from the ATIS corpus ....

.... and Buntine, 1988) and Golem (Muggleton and Feng, 1992) and top down methods from systems like Foil (Quinlan, 1990) and is able to invent new predicates in a manner analogous to Champ (Kijsirikul et al. 1992) Details of the Chill induction algorithm can be found in (Zelle and Mooney, 1993b; Zelle and Mooney, 1994). The final step in program specialization is to fold the control information back into the overly general parser. A control rule is easily incorporated into the overly general program by unifying the head of an operator clause with the head of the control rule for the clause and adding the ....

Zelle, J. M. and Mooney, R. J. (1994). Combining topdown and bottom-up methods in inductive logic programming. In Proceedings of the Eleventh International Conference on Machine Learning. New Brunswick, NJ.

No context found.

Zelle, J. M., & Mooney, R. J. (1994a). Combining top-down and bottom-up methods in inductive logic programming. In Proceedings of the Eleventh International Conference on Machine Learning New Brunswick, NJ.

....consistent clauses with the greatest coverage of positive examples. That clause is further generalized by computing the rlggs of the clause with new randomly selected positive examples, and generalization terminates when the coverage of the best consistent clause stops improving. Chillin (Zelle Mooney 1994) combines bottom up and top down techniques. The algorithm starts with a most specific definition (the complete set of positive examples) and introduces consistent generalizations that make the definition more compact. The search for consistent generalizations combines bottom up methods from ....

Zelle, J. M., and Mooney, R. J. 1994. Combining top-down and bottom-up methods in inductive logic programming. In Proceedings of the Eleventh International Conference on Machine Learning, 343--351.

....definition of the concept subgoals for which C is useful . In the blocksworld domain, such a definition is learned for each of the candidates shown in Figure 1. In this context, control rule learning can be viewed as relational concept learning. A number of systems (Quinlan 1990; Muggleton 1992; Zelle Mooney 1994) have been designed to tackle this type of learning problem. Scope employs a version of Quinlan s Foil algorithm to learn control rules through induction. The choice of a Foil like framework is motivated by a number of factors. First, the basic Foil algorithm is relatively easy to implement and ....

Zelle, J. M., and Mooney, R. J. 1994. Combining top-down and bottom-up methods in inductive logic programming. In Proceedings of the Eleventh International Conference on Machine Learning, 343--351.

....Inductive Logic Programming, a number of related pieces of research should be mentioned. The use of intensional background knowledge is the least distinguishing feature, since a number of other ILP systems also incorporate this aspect. Focl [16] mFoil [ 8] Grendel [6] Forte [20] and Chillin [25] all use intensional background to some degree in the context of a Foil like algorithm. The use of implicit negatives is significantly more novel. Bergadano et al. 2] allows the user to supply an intensional definition of negative examples that covers a large set of ground instances; however, to ....

J. M. Zelle and R. J. Mooney. Combining top-down and bottom-up methods in inductive logic programming. In Proceedings of the Eleventh International Conference on Machine Learning, pages 343--351, New Brunswick, NJ, July 1994.

No context found.

Zelle, J. M., & Mooney, R. J. (1994a). Combining top-down and bottom-up methods in inductive logic programming. In Proceedings of the Eleventh International Conference on Machine Learning New Brunswick, NJ.

No context found.

Zelle, J. M., & Mooney, R. J. (1994a). Combining top-down and bottom-up methods in inductive logic programming. In Proceedings of the Eleventh International Conference on Machine Learning, pp. 343--351 New Brunswick, NJ.

....with the greatest coverage of positive examples is selected, and that clause is further generalized by computing the rlggs of the selected clause with new randomly chosen positive examples. The generalization process stops when the coverage of the best clause no longer increases. The Chillin (Zelle and Mooney, 1994) system combines top down (general to specific) and bottomup ILP techniques. The algorithm starts with a most specific definition (the set of positive examples) and introduces generalizations which make the definition more compact. Generalizations are created by selecting pairs of clauses in the ....

Zelle, J. M. and R. J. Mooney. 1994. Combining top-down and bottom-up methods in inductive logic programming. In Proceedings of the Eleventh International Conference on Machine Learning, pages 343--351, New Brunswick, NJ, July.

....Previous experiments with learning list processing Prolog programs have employed specially constructed sets of examples that are guaranteed to be complete in some sense. However, ideally an ILP system should be able to learn such programs from random examples rather than carefully selected sets (Zelle Mooney, 1994; Cohen, 1993) Consequently, we compared systems on randomly 0 20 40 60 80 100 0 5 10 15 20 Training Examples FOIDL IFOIL FFOIL 0 20 40 60 80 100 0 5 10 15 20 Training Examples FOIDL IFOIL FFOIL last shift 0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 Training Examples FOIDL IFOIL FFOIL ....

Zelle, J. M., & Mooney, R. J. (1994). Combining top-down and bottom-up methods in inductive logic programming. In Proceedings of the Eleventh International Conference on Machine Learning, pp. 343--351 New Brunswick, NJ.

....this problem. Chill combines elements from bottom up techniques found in systems such as Cigol [19] and Golem [20] and top down methods from systems like Foil [25] and is able to invent new predicates in a manner analogous to Champ [9] Details of the Chill induction algorithm can be found in [28, 29, 27]. Given our simple example, a control rule that might be learned for the agent operator is op( X, Y,det:the] the Z] A, B) animate(Y) animate(man) animate(boy) animate(girl) Here the system has invented a new predicate to help explain the parsing decisions. Of course, the new ....

J. M. Zelle and R. J. Mooney. Combining top-down and bottom-up methods in inductive logic programming. In Proceedings of the Eleventh International Conference on Machine Learning, pages 343--351, New Brunswick, NJ, July 1994.

....its learning algorithm was inspired by ideas from three ILP systems. Golem (Muggleton Feng 1992) employs a bottom up algorithm based on the construction of relative least general generalizations, rlggs. Rules are created by computing the rlggs of randomly selected positive examples. Chillin (Zelle Mooney 1994) combines bottom up and top down techniques. The algorithm starts with a most specific definition (the set of positive examples) and introduces generalizations that compress the definition. The third system, Progol (Muggleton 1995) also combines bottom up and top down search. It constructs a most ....

Zelle, J. M., and Mooney, R. J. 1994. Combining top-down and bottom-up methods in inductive logic programming. In Proceedings of the Eleventh International Conference on Machine Learning, 343--351.

....consistent clauses with the greatest coverage of positive examples. That clause is further generalized by computing the rlggs of the clause with new randomly selected positive examples, and generalization terminates when the coverage of the best consistent clause stops improving. The Chillin [ Zelle and Mooney, 1994 ] system combines top down (general to specific) and bottomup ILP techniques. The algorithm starts with a most specific definition (the set of positive examples) and introduces generalizations which make the definition more compact. Generalizations are created by selecting pairs of clauses in the ....

J. M. Zelle and R. J. Mooney. Combining top-down and bottom-up methods in inductive logic programming. In Proceedings of the Eleventh International Conference on Machine Learning, pages 343--351, New Brunswick, NJ, July 1994.

No context found.

Zelle, J. M. and Mooney, R. J. (1994). Combining top-down and bottom-up methods in inductive logic programming. In Proceedings of the Eleventh International Conference on Machine Learning, pages 343--351, New Brunswick, NJ.

No context found.

Zelle, J. M., & Mooney, R. J. (1994a). Combining top-down and bottom-up methods in inductive logic programming. In Proceedings of the Eleventh International Conference on Machine Learning, pp. 343--351 New Brunswick, NJ.

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