Cohen, W. W. Learning Approximate Control Rules of High Utility. In Proceedings of the sixth international conference on machine learning, pages 268-276. August, 1990.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....or intractable theories, such as (Tadepalli, 1989) Alternatively, inductive ap proaches incrementally acquire correct knowledge by observing a large set of problem solving examples. These approaches strongly depend on the particular examples seen, but can also acquire simple and useful rules (Cohen, 1990, Leckie and Zukerman, 1991) This article presents a method that combines a deductive and an inductive approach, integrating three aspects of lazy learning. The results show that the combination of these three lazy components has several advantages over eager deductive approaches, such as: ....

William W. Cohen. Learning approximate control rules of high utility. In Proceed- ings of the Seventh International Conference on Machine Learning, pages 268-276, 1990.

....1993; Doorenbos, 1994#, although the combination of this result and the approach described here still needs to be analyzed. A third class of approaches to the utility problem is to use inductive learning techniques to learn simpler #or approximate# control rules with reduced match cost #Cohen, 1990; Zelle Mooney, 1993#. These approaches are on the other side of the spectrum of maintaining versus dropping #or simplifying# information for e#ciency in the learning process. BEBL Soar keeps the performance information as well as the accuracy information in learning to provide boundedness of ....

Cohen, W. W. #1990#. Learning approximate control rules of high utility.InProceedings of the Seventh International Conference on Machine Learning, pages 268#276.

....to the domain theory. The technique used in gentre to construct generalized rules by removing literals from the body of the rule (which is a development of a technique described in [3] can be thought of as the construction of partial explanations, and is similar to the one used in AxA EBL [6]. 3 A DESCRIPTION OF GENTRE Gentre 1 is a system for revision and refinement of domain theories (Gentre is described in more detail in [4] It takes as input a set of positive and negative examples of a target concept and an original (possibly incorrect) domain theory expressed in Horn clause ....

....In a second step, literals in the body of each generalization, whose instansiation are dependent on the removed literal, will also be removed. The generalized rules can also be viewed as partial explanations for the concept that they describe, or they can be viewed as approximations as in [6]. This corresponds to a top down construction of partial explanations rather than a bottom up. The unexplained training examples do not trigger the construction of partial explanations. Instead, partial explanations are generated from the explanations of successfully processed examples, which are ....

Cohen W.W. (1990). "Learning Approximate Control Rules of High Utility", the Seventh International Conference on Machine Learning, Austin, TX, pp 268--276.

....solving action model, a complete domain theory would include also a set of domain axioms that enables the proof of the universal truth of episodic explanations. problem solving examples. These approaches strongly depend on the particular examples seen, but can also acquire simple and useful rules (Cohen 1990; Leckie Zukerman 1991) HAMLET combines a deductive and an inductive approach: it generates bounded explanations from search trees that are refined inductively with more example problems. Hence, the learned knowledge becomes increasingly correct incrementally. HAMLET learns control knowledge ....

Cohen, W. W. 1990. Learning approximate control rules of high utility. In Proceedings of the Seventh International Conference on Machine Learning, 268--276.

....takes to solve problems (e.g. by pruning the search space) but if the slowdown in the matcher increases the time per step, then this can outweigh the reduction in the number of steps. This has been observed in several machine learning systems (Minton, 1988; Etzioni, 1990a; Tambe et al. 1990; Cohen, 1990; Gratch and DeJong, 1992) To avoid this slowdown, previous research on the utility problem from match cost has taken three general approaches. One approach is simply to reduce the number of rules in the system s knowledge base, by being selective about when to learn or which rules or types of ....

....for learning, rather than forgoing certain opportunities. The second general approach is to reduce the match cost of individual rules, taken one at a time. Many techniques have been developed for this (Tambe et al. 1990; Minton, 1988; Etzioni, 1990a; P erez and Etzioni, 1992; Chase et al. 1989; Cohen, 1990; Kim and Rosenbloom, 1993) This prevents just a handful of expensive rules from slowing the matcher down to a crawl; thus, the system can learn more rules before an overall slowdown results. Unfortunately, an overall slowdown can still result when a large number of individually cheap rules exact ....

Cohen, W. W. (1990). Learning approximate control rules of high utility. In Proceedings of the Seventh International Conference on Machine Learning, pages 268--276.

....DeJong Mooney, 1986) An EBL learner would analyze positive control examples in the context of a trace of the problem solver to extract the specific features which contributed to the example s success. Some systems have employed a combination of SBL and EBL to learn controlrules (Mitchell, 1984; Cohen, 1990; Zelle Mooney, 1993a) In the final phase, the learned control rules must be integrated with the initial problem solver to produce an enhanced system. Intuitively, this means adding an evaluation at each decision point to select which option (or options) should be followed up, based on the ....

....to select which option (or options) should be followed up, based on the learned control rules. 2.1. 1 Learning Search Control in Logic Programs Some recent research has investigated the learning of search control heuristics to improve the efficiency of problem solvers implemented as logic programs (Cohen, 1990; Zelle Mooney, 1993a) A logic program is expressed using the definite clause subset of first order logic. A definite clause is a disjunction of literals having exactly one unnegated literal, called the head. The negated literals comprise the clause body. Computation in logic programs is ....

[Article contains additional citation context not shown here]

Cohen, W. W. (1990). Learning approximate control rules of high utility. In Proceedings of the Seventh International Conference on Machine Learning, pp. 268--276 Austin, TX.

.... has generally focussed on learning macros (compiled rules) Mitchell et al. 1986; DeJong and Mooney, 1986; Prieditis and Mostow, 1987 ] while EBL work in planning and production systems has tended to focus on learning search control rules [ Minton, 1988; Laird et al. 1986 ] Recently, Cohen [ Cohen, 1990 ] has argued the advantages of learning search control rules for the clause selection problem in logic programming. Clause selection is the process of deciding which of several applicable clauses to use in reducing a particular subgoal in the course of a proof. Incorporating a set of accurate ....

....dramatically enhance the efficiency of an algorithm. 3 Experimental Results 3. 1 Experimental Design The Dolphin system has been evaluated on five problem domains: Two generate and test programs, naivesort and N queens, and three standard EBL problems LEX, RW, and BW borrowed from [ Cohen, 1990 ] The N queens problem is adapted from a Prolog program given in [ Bratko, 1990 ] The problem is to find a placement of N queens on an NxN chessboard such that no queen is attacking another. The program implements a generate and test strategy where a configuration is represented by a ....

[Article contains additional citation context not shown here]

W. W. Cohen. Learning approximate control rules of high utility. In Proceedings of the Seventh International Conference on Machine Learning, pages 268--276, Austin, TX, June 1990.

.... and CINDI [ Callan and Utgoff, 1991 ] Also, the INTSUM procedure of META Dendral [ Buchanan and Mitchell, 1978 ] uses a half order domain theory as input to inductive learning; IOE [ Flann and Dietterich, 1989 ] uses both EBL and inductive methods in a single process to develop a concept; and Cohen [ 1990 ] does induction over generalizations of EBG rules. Each of these approaches studies how training data can be supplemented through the use of a 39 domain theory. 2.1.2 Analogy The idea of transfer presented here has strong conceptual ties to the idea of analogy that s studied in symbolic ....

William Cohen. Learning approximate control rules of high utility. In Proceedings of the Seventh International Conference on Machine Learning, pages 268--276, Austin, Texas, July 1990.

....rather than improve overall planning This research was supported by the NASA Graduate Student Researchers Program, grant number NGT 51332. performance (Minton 1989) By incorporating induction to learn simpler, approximate control rules, we can greatly improve the utility of acquired knowledge (Cohen 1990). Scope (Search Control Optimization of Planning through Experience) integrates explanation based generalization (EBG) Mitchell et al. 1986; DeJong Mooney, 1986) with techniques from inductive logic programming (ILP) Quinlan 1990; Muggleton 1992) to learn high utility rules that can ....

....inappropriate refinement will immediately fail. The Prolog programming language provides an excellent framework for learning control rules. Search algorithms can be implemented in Prolog in such a way that allows control information to be easily incorporated in the form of clause selection rules (Cohen 1990). These rules help avoid inappropriate clause applications, thereby reducing backtracking. A version of the UCPOP partial order planning algorithm has been implemented as a Prolog program. 1 Planning decision points are represented in this program as clause selection problems (i.e. each ....

[Article contains additional citation context not shown here]

Cohen, W. W. 1990. Learning approximate control rules of high utility. In Proc. of ML-90, 268--276.

....refinement will immediately fail. Scope is implemented in Prolog, which provides an excellent framework for learning control rules. Search algorithms can be implemented in Prolog in such a way that allows control information to be easily incorporated in the form of clause selection rules [ Cohen, 1990 ] For its base planner, Scope uses a version of the UCPOP partial order planning algorithm which has been reimplemented in Prolog. 2 Planning decision points are represented in the planner as clause selection problems (i.e. 2 The main difference between our planner and UCPOP is New ....

W. W. Cohen. Learning approximate control rules of high utility. In Proceedings of the Seventh International Conference on Machine Learning, pages 268--276, Austin, TX, June 1990.

....or intractable theories, such as (Tadepalli, 1989) Alternatively, inductive approaches incrementally acquire correct knowledge by observing a large set of problem solving examples. These approaches strongly depend on the particular examples seen, but can also acquire simple and useful rules (Cohen, 1990; Leckie and Zukerman, 1991) This article presents a method that combines a deductive and an inductive approach, integrating three aspects of lazy learning. The results show that the combination of these three lazy components has several advantages over eager deductive approaches, such as: ....

Cohen, W. W. (1990). Learning approximate control rules of high utility. In Proceedings of the Seventh International Workshop on Machine Learning, pages 268--276, Austin, TX. Morgan Kaufmann.

....for planning systems. Our learning system Scope uses a unique combination of machine learning techniques to acquire effective search control rules for planning. By incorporating induction with EBL to learn simple, approximate control rules, we can greatly improve the utility of acquired knowledge (Cohen, 1990; Leckie Zuckerman, 1993) Specifically, Scope (Search Control Optimization of Planning through Experience) integrates explanation based generalization (EBG) Mitchell et al. 1986; DeJong Mooney, 1986) with techniques from inductive logic programming (ILP) Quinlan, 1990; Muggleton, 1992; ....

.... well on more complicated tasks such as learning properties of organic molecules (Muggleton et al. 1992) and predicting the past tense of English verbs (Mooney Califf, 1995) More recently, it has been argued that ILP techniques can also be a useful tool for acquiring control information (Cohen, 1990). Many different problem solving strategies can be easily coded as Prolog programs and learning mechanisms are also easily implemented in this framework. Logic programs have long been recognized as a good platform for EBL techniques since the notion of explanation can be equated with the ....

[Article contains additional citation context not shown here]

Cohen, W. W. (1990). Learning approximate control rules of high utility. In Proceedings of the Seventh International Conference on Machine Learning, pp. 268--276 Austin, TX.

....an oracle. Efficient oracle free methods for predicate invention are needed to revise programs that require additional recursive subroutines. 2 The terms static and dynamic are borrowed from [ Murray, 1988 ] We are also developing techniques for learning search heuristics [Mitchell, 1984; Cohen, 1990] to improve the efficiency of logic programs. Meta rules for when to use a particular clause can be empirically learned using sample calls for which the clause ultimately failed or succeeded in leading to a final solution. Such examples can be extracted from the search conducted during the ....

....of logic programs. Meta rules for when to use a particular clause can be empirically learned using sample calls for which the clause ultimately failed or succeeded in leading to a final solution. Such examples can be extracted from the search conducted during the execution of a logic program [Cohen, 1990]. Existing ILP systems should be useful for learning search heuristics from these examples. As an example, consider the following exponential time sorting program: naivesort(X,Y) permutation(X,Y) ordered(Y) permutation( permutation( X Xs] Ys) permutation(Xs,Ys1) ....

W. W. Cohen. Learning approximate control rules of high utility. In Proceedings of the Seventh International Conference on Machine Learning, pages 268--276, Austin, TX, June 1990.

....and dl [1] be a decision list. By D [i] and dl [i] we denote the values of internal variables D and dl at the beginning of the ith iteration of while loop section with inputs, T , D [1] l max and dl [1] Further, a ( possibly infinite ) sequence of tuples (D [1] dl [1] D [2] ; dl [2] is called a computation process of the while loop section with inputs, T , D [1] l max and dl [1] It suffices to show the next claim. Claim : The computation process (D [1] dl [1] D [2] dl [2] is finite for any inputs of T , D [1] l max and ....

....be a decision list. By D [i] and dl [i] we denote the values of internal variables D and dl at the beginning of the ith iteration of while loop section with inputs, T , D [1] l max and dl [1] Further, a ( possibly infinite ) sequence of tuples (D [1] dl [1] D [2] dl [2] ) is called a computation process of the while loop section with inputs, T , D [1] l max and dl [1] It suffices to show the next claim. Claim : The computation process (D [1] dl [1] D [2] dl [2] is finite for any inputs of T , D [1] l max and dl [1] ....

[Article contains additional citation context not shown here]

W. W. Cohen. : "Learning Approximate Control Rules of High Utility", Proc. of International Conference on Machine Learning'90, pp.268-276 (1990).

.... Unfortunately, standard EBL can frequently produce complex, overly specific control rules that decrease rather than improve overall planning performance (Minton 1988) By incorporating induction to learn simpler, approximate control rules, we can greatly improve the utility of acquired knowledge (Cohen 1990; Leckie Zuckerman 1993) In this paper, we describe Scope, a system that uses a unique combination of machine learning techniques to acquire effective search control rules for a partial order planner. This research was supported by the NASA Graduate Student Researchers Program, grant number ....

....depending on the accuracy of the learned rule. The Prolog programming language provides an excellent framework for learning control rules. Search algorithms can be implemented in Prolog in such a way that allows control information to be easily incorporated in the form of clause selection rules (Cohen 1990). These rules help avoid inappropriate clause applications, thereby reducing backtracking. A version of the UCPOP partial order planning algorithm has been implemented as a Prolog program. 2 Planning decision points are represented in this program as clause selection problems (i.e. each ....

[Article contains additional citation context not shown here]

Cohen, W. W. 1990. Learning approximate control rules of high utility. In Proceedings of the Seventh International Conference on Machine Learning, 268-- 276.

....refinement will immediately fail. Scope is implemented in Prolog, which provides an excellent framework for learning control rules. Search algorithms can be implemented in Prolog in such a way that allows control information to be easily incorporated in the form of clause selection rules [ Cohen, 1990 ] For its base planner, Scope uses a version of the UCPOP partial order planning algorithm which has been reim New Planning Program Training Examples Planning Program Selection Decisions Generalized Proof Trees Selection Rules Program Specialization Example Analysis Control Induction Rule ....

W. W. Cohen. Learning approximate control rules of high utility. In Proceedings of the Seventh International Conference on Machine Learning, pages 268--276, Austin, TX, June 1990.

.... can communicate in two directions: a theory can be used to focus the use of the training data, and inductive learning can generate new rules in the language of the domain theory ( Pazzani, 1989, Mooney and Ourston, 1989, Danyluk, 1989, Mooney and Ourston, 1989, Buchanan and Mitchell, 1978, Cohen, 1990 ] Several systems also use domain knowledge to create new features for inductive learning ( Utgoff, 1986, Fawcett and Utgoff, 1991, Callan and Utgoff, 1991 ] Each of these approaches attempts to address the question of how to supplement training data with extra information. They can ....

William Cohen. Learning approximate control rules of high utility. In Proceedings of the Seventh International Conference on Machine Learning, pages 268--276, Austin, Texas, July 1990.

....a third approach: specializing an existing program by learning control rules that restrict the application of specific clauses. Induction of control rules has a fairly long history in learning and problem solving (Mitchell, 1983; Langley, 1985) and more recent work has applied ILP to this task (Cohen, 1990; Leckie Zuckerman, 1993; Zelle Mooney, 1993a; Estlin Mooney, 1996) These systems focus on learning control rules that improve the efficiency of an existing program, such as transforming an O(n ) naive sorting program into an O(n 2 ) insertion sort (Zelle Mooney, 1993a) Chill ....

Cohen, W. W. (1990). Learning approximate control rules of high utility. In Proceedings of the Seventh International Conference on Machine Learning, pp.

....the swamping problem and the expensive chunks problem. 4.1.1. SWAMPING Swamping was first noticed in PRODIGY EBL (MINTON 1988) Swamping occurs when a system learns a large number of rules whose combined match time slows problem solving time down more than individual control rules speed it up (COHEN 1990). In other words, the calculated speedup F provided by the database is outweighed by the slowdown S(O K ) of the database. Each individual rule may be useful and easy to match, but the overhead of matching the rule set outweighs any performance benefit provided by the individual rules. Swamping ....

Cohen, W.W. "Learning approximate control rules of high utility." In Machine Learning: Proceedings of the Seventh International Conference, 1990.

....extract sufficient conditions for future applications of the operator. The learning of search control knowledge has been previously investigated primarily in the context of STRIPS like planners (Minton, 1988) and forward chaining production systems (Langley, 1985; Laird et al. 1986) Recently, Cohen (1990) has argued some advantages of extending the search control learning framework into the domain of logic programming. This proposal embraces such an approach and offers a general framework for learning searchcontrol heuristics in logic programs. The motivation for this research is three fold. ....

....is one that computes a (partial) function; that is, it finds at most a single solution for each unique instantiation of its input arguments. A nondeterministic program computes a relation, admitting multiple solutions. Search control in a logic program can be viewed as a clause selection problem (Cohen, 1990). Clause selection is the process of deciding which of several applicable program clauses should be used to reduce a particular subgoal during the course of a proof. If program clauses are always applied appropriately, the program executes deterministically (without backtracking) and produces only ....

[Article contains additional citation context not shown here]

Cohen, W. W. (1990). Learning approximate control rules of high utility. In Proceedings of the Seventh International Conference on Machine Learning, pages 268--276. Austin, TX.

....Sample problems are used to generate a separate set of control examples describing appropriate and inappropriate contexts for applying operators in the domain theory. A combination of ILP and EBL methods are then used to learn control rules for deciding when to apply the existing operators, e.g. [4, 55]. Control learning is traditionally used to improve the efficiency of a problem solver as a form of speedup learning [46] however, it can also be used to improve accuracy [56] This paper presents a review of recent research that integrates ILP and EBL methods. We intentionally focus on ILP work ....

....the context of logic programming. 4. 2 Controlling Search in Logic Programs Although most EBL research in logic programming has generally focussed on learning macros, 25, 10, 36] Recent work has shown the utility of learning explicit search control rules within a logic programming framework [4, 55, 56]. The execution of a logic program can be viewed as a problem solving process with a search strategy based on resolution theorem proving. A program executes by finding a constructive proof of a partially instantiated goal given as input. Prolog provides a particular implementation of logic ....

[Article contains additional citation context not shown here]

W. W. Cohen. Learning approximate control rules of high utility. In Proceedings of the Seventh International Conference on Machine Learning, pages 268--276, Austin, TX, June 1990.

.... abstracted explanations, as was done in Experiment 2. A final application of A EBL (of special interest to us because it was one of the original motivation for pursuing this research) is learning control rules for search programs. Some initial results in this area have been reported in [Cohen, 1990c] 7 Conclusion A much investigated research topic in machine learning is using prior knowledge of a learning problem to improve the performance of similarity based learning techniques. One approach to this problem is to attempt to extend explanation based learning (EBL) methods to imperfect ....

William W. Cohen. Learning approximate control rules of high utility. In Proceedings of the Seventh International Conference on Machine Learning, Austin, Texas, 1990. Morgan Kaufmann.

No context found.

Cohen, W. W. Learning Approximate Control Rules of High Utility. In Proceedings of the sixth international conference on machine learning, pages 268-276. August, 1990.

No context found.

Cohen, W. (1990). Learning approximate control rules of high utility. Proc. the Seventh International Conference on Machine Learning, 268-- 276, Morgan Kaufmann, San Mateo.

No context found.

Wadsworth. Cohen, W. W. 1990. Learning approximate control rules of high utility. In Proceedings of the Seventh International Conference on Machine Learning. 268-- 276.

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