Ourston, D.: Using Explanation-Based and Empirical Methods in Theory Revision. Ph.D. thesis. The University of Texas, Austin, Texas, USA (1991)  Home/Search   Document Not in Database   Summary   Related Articles   Check

....striking aspect of these numbers is that all measures of theory size are relatively stable with respect to training set size. Naturally,the exact values are to a large degree an artifact of the inductive learning component used. In contrast, for EITHER, theory size increases with training set size [Ourston91]. Forexample, for 20 training examples the output theory size (clauses plus literals) is 78, while for 80 training examples, the output theory size is 106. Unfortunately,making direct comparisons with KBANN or RAPTURE is difficult. In the case of KBANN and RAPTURE, which allownumerical rules, ....

D. Ourston, Using Explanation-Based and Empirical Methods in Theory Revision,Ph.D. Thesis, University of Texas at Austin, Austin, TX (August 1991).

....Rapture also does not fully connect between network layers, allowing less freedom for the network to modify the initial domain rules during training. 8.2. Symbolic Theory Revision A number of other systems have been developed which take a much more directed approach to theory revision. Either (Ourston, 1991) was developed to revise domain theories in predicate logic form. The Either algorithm can modify theories by adding or deleting either rules or their antecedents. Neither (Baffes Mooney, 1993) was developed as an extension to Either. It speeds up Either s algorithm and introduces the concept of ....

Ourston, D. (1991). Using Explanation Based and Empirical Methods on Theory Revision. PhD thesis, University of Texas, Austin.

....Rapture also does not fully connect between network layers, allowing less freedom for the network 17 to modify the initial domain rules during training. 3.3 Symbolic Theory Revision A number of other systems have been developed which take a much more directed approach to theory revision. Either [Our91] was developed to revise domain theories in predicate logic form. The Either algorithm can modify theories by adding or deleting either rules or their antecedents. Neither [BM93] was developed as an extension to Either. It speeds up Either s algorithm and introduces the concept of M of N rules to ....

D. Ourston. Using Explanation Based and Empirical Methods on Theory Revision. PhD thesis, University of Texas, Austin, 1991.

....scratch (Ourston and Mooney, 1994; Towell and Shavlik, 1993) chapter.tex; 6 07 1998; 11:53; p.4 5 3.2. Theory Refinement Algorithms and Systems Several theory refinement systems use abduction on individual examples to locate faults in a theory and suggest repairs (Ourston and Mooney, 1990; Ourston, 1991; Ourston and Mooney, 1994; Wogulis and Pazzani, 1993; Wogulis, 1994; Baffes and Mooney, 1993; Baffes, 1994; Baffes and Mooney, 1996; Brunk, 1996) The ways in which various forms of logical abduction can be used in revising theories is also discussed and reviewed by Dimopoulos and Kakas (1996) ....

....as the number of positive examples that it covers. For example, the more negative examples that are generated when the literals corresponding to an assumption set are deleted, the more complex the resulting repair is likely to be. The Either (Ourston and Mooney, 1990; Ourston and Mooney, 1994; Ourston, 1991) and Neither (Baffes and Mooney, 1993; Baffes, 1994) theory refinement systems allow multiple assumptions in order to prove an example, preferring more specific assumptions, i.e. they employ most specific abduction (Cox and Pietrzykowski, 1987) Audrey (Wogulis, 1991) Audrey II (Wogulis and ....

Ourston, D.: 1991, `Using Explanation-Based and Empirical Methods in Theory Revision'. Ph.D. thesis, Department of Computer Sciences, University of Texas, Austin, TX. Also appears as Artificial Intelligence Laboratory Technical Report AI 91-164.

.... (the bars labeled Network ) Also included in Figure 10 is the accuracy of the Either system, an all symbolic method for the empirical adaptation of rules which has been tested using the promoter dataset (Ourston and Mooney, 1990) The numbers for Either are derived from Ourston s thesis (Ourston, 1991); they reflect a slightly different testing method. Either has not been tested on the splice junction problem. Recall the initial rule sets for promoter recognition and splice junction determination correctly categorized 50 and 61 , respectively, of the examples. Hence, each of the systems ....

....that match the parenthesized sequence. rewritten in a form very similar to one used in the biological community (Stormo, 1990) namely weight matrices. The major pattern in the extracted rules is that Kbann learned to disregard conformation. The conformation rules are also dropped by Either (Ourston, 1991), which suggests that dropping these rules is not an artifact of Kbann but rather that DNA bases outside the minus35 and minus10 regions are less important than the conformation hypothesis (Koudelka et al. 1987) suggests. Hence, machine learning methods can provide valuable evidence confirming ....

Ourston, D., Using Explanation-Based and Empirical Methods in Theory Revision, PhD thesis, Deptartment of Computer Sciences, University of Texas, Austin, TX, 1991.

.... them slightly more than two years to complete the task, and they admitted that the resulting rules were still not perfect, though they were able to correctly classified returns over 90 of the time (Davis, 1995) Out of an attempt to alleviate this bottleneck, researchers (Ginsberg, 1988; Mooney Ourston, 1991; Towell, Shavlik, Noordewier, 1990; Pazzani Kibler, 1992) have sought ways of automating the process of fine tuning a rule base. Throughout this document, I refer to this process as rule base revision, though other authors have called it theory revision (Mitchell, Keller, Kedar Cabelli, ....

....size of the rule base, makes it more difficult to revise, as well as less understandable for the human. Flexible matching methods, which have been used successfully in inductive rule learning (Michalski Chilausky, 1980; Michalksi, Mozetic, Hong, Lavrac, 1986) and rule base revision (Mooney Ourston, 1991), are one way of dealing with this problem. In this approach, a score is calculated measuring how well an example matches each symbolic rule, and the rule with the greatest score is invoked. Another approach is to use M of N rules (Towell Shavlik, 1991; Ginsberg, 1988) which fire if at least M ....

[Article contains additional citation context not shown here]

Ourston, D. (1991). Using Explanation-Based and Empirical Methods in Theory Revision. Ph.D.

....a significantly more accurate theory with minor revisions that are easy to understand. 2 Theory Revision Algorithm 2. 1 The Either Algorithm The original Either theory refinement algorithm has been presented in various levels of detail in [ Ourston and Mooney, 1990; Ourston and Mooney, in press; Ourston, 1991 ] It was designed to repair propositional Horn clause theories that are either overly general or In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI 93) Chambery, France, pp. 1135 1140, 1993. overly specific or both. An overly general theory is one that ....

....Either retracts and generalizes existing antecedents and learns new rules to fix these problems. Unlike other theory revision systems that perform hill climbing (and are therefore subject to local maxima) Either is guaranteed to fix any arbitrarily incorrect propositional Horn clause theory [ Ourston, 1991 ] Either Main Loop Compute all repairs for each example While some examples remain uncovered Add best repair to cover set Remove examples covered by repair end Apply repairs in cover set to theory Neither Main Loop While some examples remain Compute a single repair for each example Apply ....

D. Ourston. Using Explanation-Based and Empirical Methods in Theory Revision. PhD thesis, Department of Computer Sciences, University of Texas, Austin, TX, August 1991.

....Experimental results on the audiology domain show that using referees can help obtain higher accuracies than those obtained by any of the individual prediction models. In this paper, two currently active research programs in machine learning, theory revision (Flann Dietterich 1989) Mooney 1993) (Ourston 1991) (Baffes Mooney 1993) Richards Mooney 1995) Towell, Shavlik, Noordewier 1990) R. S. Michalski 1993) Cohen 1992) Bergadano Giordana 1988) and bias selection (Merz 1995) Ho, Hull, Srihari 1994) Brodley 1993) Schaffer 1993) are viewed from a single perspective. Theory revision ....

Ourston, D. 1991. Using Explanation-Based and Empirical Methods in Theory Revision. Ph.D. Dissertation, University of Texas, Austin, TX.

....Noordewier and J. Shavlik to the UCI repository (Murphy Aha, 1992) have become popular for testing systems that integrate empirical and analytical learning (Hirsh Japkowicz, 1994; Koppel, Feldman, Segre, 1994b; Mahoney Mooney, 1994, 1993; Norton, 1994; Opitz Shavlik, 1994; Ortega, 1994; Ourston, 1991; Towell, Shavlik, Noordewier, 1990; Shavlik, Towell, Noordewier, 1992) The original domain theory, as usually interpreted, is overly specific in that it classifies all of the promoter sequences in the database as negative instances. Since the database consists of 53 positive instances and 53 ....

....ways to loosen the conditions of the domain theory. First, the conformation condition has very weak biological support. This was implied by the initial KBANN experiments (Shavlik et al. 1992) where none of the learned rules referenced the conformation conditions. In addition, the EITHER system (Ourston, 1991) eliminated rules involving conformation altogether from the domain theory. Eliminating conformation was also supported by a domain expert (Ourston, 1991) The second piece of domain knowledge is that the concepts in this domain tend to take the form of M of N concepts. Some of the final rules ....

[Article contains additional citation context not shown here]

Ourston, D. (1991). Using Explanation-Based and Empirical Methods in Theory Revision.

....do not agree with (cannot be explained by) the model, but differs from them in that we try to justify each rule. Our focus is on inducing rules that are explainable, rather than on repairing a faulty or incomplete domain theory. In that sense, our work is different from theory revision as done in [17] and [22] It has to be emphasized here, however, that a recent extension of our approach can be used as a theory revision technique when there is overwhelming evidence in the training data and the data persistently conflicts with the existing theory ( 13] This work is organized as follows. To ....

Ourston, D. (1991). Using Explanation-based and Empirical Methods in Theory Revision. Ph.D. Thesis. University of Texas. Austin, TX.

....upon the language used to represent student knowledge. 3 Theory Refinement Algorithm For its student modeling component, Assert uses a propositional Horn clause theory refinement system based upon the Either theory refinement algorithm (Ourston and Mooney, 1990; Ourston and Mooney, in press; Ourston, 1991). Either was selected because it is the most complete theory refinement system available. It can generalize or specialize a theory, and is guaranteed to produce a set of refinements which are consistent with the input examples. Unfortunately, Either s worst case run time is exponential in the size ....

....Either retracts and generalizes existing antecedents and learns new rules to fix these problems. Unlike other theory refinement systems that perform hill climbing (and are therefore subject to local maxima) Either is guaranteed to fix any arbitrarily incorrect propositional Horn clause theory (Ourston, 1991). As an example of how Either repairs a theory, refer to Figure 1. This illustration depicts Compute all minimal repairs for each example While some examples remain uncovered Add best repair to covering set Remove examples covered by repair end Apply repairs in covering set to theory Figure 2: ....

Ourston, D. (1991). Using Explanation-Based and Empirical Methods in Theory Revision.

....b c a b d b e c f g o p q c g Figure 1: Partial proofs for unprovable positive example. Unprovable antecedents are shown with dotted lines. 2 Theory Revision Algorithm 2. 1 The Either Algorithm The original Either theory refinement algorithm has been presented in various levels of detail in [11, 12, 10]. It was designed to revise propositional Horn clause theories. For Either, a theory is a set of propositional Horn clause rules such as those shown in the top half of Figure 1. Each theory is assumed to function as a classification system whereby examples are labeled as belonging to one of a ....

....Rules Generalized Rules Deleted Rules New Rules Specialized Rules Figure 2: Block diagram of Either. Unlike other theory revision systems that perform hillclimbing (and are therefore subject to local maxima) Either is guaranteed to fix any arbitrarily incorrect propositional Horn clause theory [10]. The basic algorithm used by Either for both generalization and specialization is shown in the top half of Figure 3. There are three steps. First, all possible repairs for each failing example are computed. Next, Either enters a loop to compute a subset of these repairs that can be applied to ....

D. Ourston. Using Explanation-Based and Empirical Methods in Theory Revision. PhD thesis, University of Texas, Austin, TX, August 1991. Also appears as Artificial Intelligence Laboratory Technical Report AI 91-164.

.... by an expert results in more accurate results than inducing a knowledge base from scratch [ Ourston and Mooney, 1994; Towell and Shavlik, 1993 ] Several theory refinement systems use abduction on individual examples to locate faults in a theory and suggest repairs [ Ourston and Mooney, 1990; Ourston, 1991; Ourston and Mooney, 1994; Wogulis and Pazzani, 1993; Wogulis, 1994; Baffes and Mooney, 1993; Baffes, 1994; Baffes and Mooney, 1996; Brunk, 1996 ] Each of these systems use abduction in a slightly different way, but the following discussion summarizes the basic approach. For each individual ....

....is the simplest clause that covers both of the positive examples without covering either of the negatives. Note that although the alternative, equally simple clause Q(X) W(X) covers both positive examples, it also covers the negative example Q(d) The Either [ Ourston and Mooney, 1990; 1994; Ourston, 1991 ] and Neither [ Baffes and Mooney, 1993; Baffes, 1994 ] theory refinement systems allow multiple assumptions in order to prove an example, preferring more specific assumptions, i.e. they employ most specific abduction [ Cox and Pietrzykowski, 1987 ] Audrey [ Wogulis, 1991 ] Audrey II [ ....

D. Ourston. Using ExplanationBased and Empirical Methods in Theory Revision. PhD thesis, University of Texas, Austin, TX, August 1991. Also appears as Artificial Intelligence Laboratory Technical Report AI 91-164.

....Towell and Shavlik, 1991 ] can easily be made incremental. This paper presents empirical results on an incremental batch [ Clearwater et al. 1989 ] version of Either, a revision system for refining arbitrarily incorrect propositional Horn clause theories [ Ourston and Mooney, 1990; Mooney and Ourston, 1991b ] After processing a small batch of training examples, the resulting revised theory is fed back as the input theory for processing the next batch. In the limit, the system can be made completely incremental by setting the batch size to one. Using a theory revision system in incremental batch ....

....antecedents, and missing antecedents. In most trials, one hundred random training examples are sufficient to produce a fully corrected animal theory. Refinement Algorithm Either s theory refinement algorithm is presented in various levels of detail in [ Ourston and Mooney, 1990; Mooney and Ourston, 1991b; Ourston, 1991 ] It was designed to correct theories that are either overly general or overly specific or both. An overly general theory is one that causes an example (called a failing negative) to be classified in categories other than its own. Either specializes existing antecedents, adds ....

[Article contains additional citation context not shown here]

D. Ourston. Using Explanation-Based and Empirical Methods in Theory Revision. PhD thesis, University of Texas, Austin, TX, August 1991.

No context found.

Ourston, D.: Using Explanation-Based and Empirical Methods in Theory Revision. Ph.D. thesis. The University of Texas, Austin, Texas, USA (1991)

....hypothesis given sufficient number of examples and evaluates the computational complexity of the revision algorithm. 6. 1 Convergence Results The Either algorithm has been analyzed within the context of PAC (Probably Approximately Correct) learnability theory [47] with details presented in [32]. In summary, we can apply the following result (Theorem 4.4) from [17] Let H be a hypothesis space and L be a learning algorithm that uses H consistently (see definition below) For any 0 ffl; ffi 1, given m (ln (1=ffi) ln jHj) ffl; 10) independent random examples of any target concept ....

....log s) O(s log s) Table 1: Complexity Results Since Either s hypothesis space is propositional Horn clause theories, if there are n observable binary features the the size of the hypothesis space is 2 2 n . A detailed argument that Either uses this hypothesis space consistently is given in [32]. The argument hinges on the fact that the algorithm finds a cover of rules for fixing all of the failing examples for both generalization and specialization and that these two processes do not interfere with each other. Consequently, the above theorem applies and equation 10 indicates that: m ....

[Article contains additional citation context not shown here]

D. Ourston. Using Explanation-Based and Empirical Methods in Theory Revision. PhD thesis, University of Texas, Austin, TX, August 1991.

....; 1) 8C i (C i 6= CE ) T [ E 6j= C i ) 2) where T represents the corrected theory, E represents the conjunction of facts describing any example in the training set, CE is the correct category of the example, and C i is any arbitrary category. A proof of this consistency property is given in [Ourston, 1991]. 2.2 Types of theory errors Figure 4 shows a taxonomy for incorrect propositional Horn clause theories. At the top level, theories can be incorrect because they are either overly general or overly specific. An overly general theory entails category OVERLY SPECIFIC OVERLY GENERAL MISSING RULE ....

....reasoning methods are developed. The following sections describe each of Either s components and their interactions in details. The discussion focuses on the basic multistrategy approach employed in Either. Recent enhancements to the system are discussed in [Ourston and Mooney, 1991; Mooney and Ourston, 1991a; Mooney and Ourston, 1991b] and a complete description is given in [Ourston, 1991] 3 THE DEDUCTIVE COMPONENT The deductive component of Either is a standard backward chaining, Hornclause theorem prover similar to Prolog. Our particular implementation is based on the deductive retrieval ....

[Article contains additional citation context not shown here]

D. Ourston. Using Explanation-Based and Empirical Methods in Theory Revision. PhD thesis, University of Texas, Austin, TX, August 1991.

No context found.

D. Ourston. Using Explanation-Based and Empirical Methods in Theory Revision. PhD thesis, University of Texas, Austin, TX, 1991.

No context found.

D. Ourston. Using Explanation Based and Empirical Methods on Theory Revision. PhD thesis, University of Texas, Austin, 1991.

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