MIME-Version: 1.0
Server: CERN/3.0
Date: Tuesday, 07-Jan-97 15:56:22 GMT
Content-Type: text/html
Content-Length: 32358
Last-Modified: Wednesday, 28-Aug-96 15:56:47 GMT

<title>Theory Refinement</title><h1>Theory Refinement</h1>To view a paper, click on the open book image. <br> <br><ol><! ===========================================================================><a name="rapture-dissertation-96.ps.Z"</a><b><li>Combining Symbolic and Connectionist Learning Methods to RefineCertainty-Factor Rule-Bases<br></b>J. Jeffrey Mahoney<br>Ph.D. Thesis, Department of Computer Sciences, University of Texas at Austin, May, 1996.<p><blockquote>This research describes the system RAPTURE, which is designed torevise rule bases expressed in certainty-factor format.  Recentstudies have shown that learning is facilitated when biased withdomain-specific expertise, and have also shown that many real-worlddomains require some form of probabilistic or uncertain reasoning inorder to successfully represent target concepts. RAPTURE was designedto take advantage of both of these results. <p>Beginning with a set of certainty-factor rules, along withaccurately-labelled training examples, RAPTURE makes use of bothsymbolic and connectionist learning techniques for revising the rules,in order that they correctly classify all of the training examples. Amodified version of backpropagation is used to adjust the certaintyfactors of the rules, ID3's information-gain heuristic is used to addnew rules, and the Upstart algorithm is used to create new hiddenterms in the rule base. <p>Results on refining four real-world rule bases are presented thatdemonstrate the effectiveness of this combined approach.  Two of theserule bases were designed to identify particular areas in strands ofDNA, one is for identifying infectious diseases, and the fourthattempts to diagnose soybean diseases.  The results of RAPTURE arecompared with those of backpropagation, C4.5, KBANN, and otherlearning systems.  RAPTURE generally produces sets of rules that aremore accurate that these other systems, often creating smaller sets ofrules and using less training time. <p></blockquote><!WA0><a href="file://ftp.cs.utexas.edu/pub/mooney/papers/rapture-dissertation-96.ps.Z"><!WA1><img align=top src="http://www.cs.utexas.edu/users/ml/paper.xbm"></a><p><! ===========================================================================><a name="banner-proposal-95.ps.Z"</a><b><li> Refinement of Bayesian Networks by Combining Connectionist andSymbolic Techniques <br></b>Sowmya Ramanchandran<br>Ph.D. proposal, Department of Computer Sciences, University of Texasat Austin, 1995. <p><blockquote>Bayesian networks provide a mathematically sound formalism forrepresenting and reasoning with uncertain knowledge and are as suchwidely used. However, acquiring and capturing knowledge in thisframework is difficult. There is a growing interest in formulatingtechniques for learning Bayesian networks inductively. While theproblem of learning a Bayesian network, given complete data, has beenexplored in some depth, the problem of learning networks withunobserved causes is still open. In this proposal, we view thisproblem from the perspective of theory revision and present a novelapproach which adapts techniques developed for revising theories insymbolic and connectionist representations.  Thus, we assume that thelearner is given an initial approximate network (usually obtained froma expert). Our technique inductively revises the network to fit thedata better.  Our proposed system has two components: one componentrevises the parameters of a Bayesian network of known structure, andthe other component revises the structure of the network. Thecomponent for parameter revision maps the given Bayesian network intoa multi-layer feedforward neural network, with the parameters mappedto weights in the neural network, and uses standard backpropagationtechniques to learn the weights. The structure revision component usesqualitative analysis to suggest revisions to the network when it failsto predict the data accurately. The first component has beenimplemented and we will present results from experiments on real worldclassification problems which show our technique to be effective.  Wewill also discuss our proposed structure revision algorithm, our plansfor experiments to evaluate the system, as well as some extensions tothe system.</blockquote><!WA2><a href="file://ftp.cs.utexas.edu/pub/mooney/papers/banner-proposal-95.ps.Z"><!WA3><img align=top src="http://www.cs.utexas.edu/users/ml/paper.xbm"></a><p><! ===========================================================================><a name="assert-aaai-96.ps.Z"</a><b><li>A Novel Application of Theory Refinement to Student Modeling<br></b>Paul Baffes and Raymond J. Mooney<br><cite>Proceedings of the Thirteenth National Conference on Aritificial Intelligence</cite>,pp. 403-408, Portland, OR, August, 1996. (AAAI-96)<p><blockquote>Theory refinement systems developed in machine learning automaticallymodify a knowledge base to render it consistent with a set ofclassified training examples. We illustrate a novel application ofthese techniques to the problem of constructing a student model for anintelligent tutoring system (ITS). Our approach is implemented in anITS authoring system called Assert which uses theory refinement tointroduce errors into an initially correct knowledge base so that itmodels incorrect student behavior. The efficacy of the approach hasbeen demonstrated by evaluating a tutor developed with Assert with 75students tested on a classification task covering concepts from anintroductory course on the C++ programming language. The systemproduced reasonably accurate models and students who received feedbackbased on these models performed significantly better on a post testthan students who received simple reteaching.</blockquote><!WA4><a href="file://ftp.cs.utexas.edu/pub/mooney/papers/assert-aaai-96.ps.Z"><!WA5><img align=top src="http://www.cs.utexas.edu/users/ml/paper.xbm"></a><p><! ===========================================================================><a name="assert-jaied-95.ps.Z"</a><b><li> Refinement-Based Student Modeling and Automated Bug Library Construction<br></b>Paul Baffes and Raymond Mooney<br><cite>Journal of Artificial Intelligence in Education</cite>, 7, 1(1996), pp. 75-116.<p><blockquote>A critical component of model-based intelligent tutoring sytems is amechanism for capturing the conceptual state of the student, whichenables the system to tailor its feedback to suit individual strengthsand weaknesses.  To be useful such a modeling technique must be<em>practical</em>, in the sense that models are easy to construct, and<em>effective</em>, in the sense that using the model actually impacts studentlearning.  This research presents a new student modeling techniquewhich can automatically capture novel student errors using onlycorrect domain knowledge, and can automatically compile trends acrossmultiple student models.  This approach has been implemented as acomputer program, ASSERT, using a machine learning technique called<em>theory refinement</em>, which is a method for automatically revising aknowledge base to be consistent with a set of examples.  Using aknowledge base that correctly defines a domain and examples of astudent's behavior in that domain, ASSERT models student errors bycollecting any refinements to the correct knowledege base which arenecessary to account for the student's behavior.  The efficacy of theapproach has been demonstrated by evaluating ASSERT using 100 studentstested on a classification task covering concepts from an introductorycourse on the C++ programming language.  Students who receivedfeedback based on the models automatically generated by ASSERTperformed significantly better on a post test than students whoreceived simple teaching.</blockquote><!WA6><a href="file://ftp.cs.utexas.edu/pub/mooney/papers/assert-jaied-95.ps.Z"><!WA7><img align=top src="http://www.cs.utexas.edu/users/ml/paper.xbm"></a><p><! ===========================================================================><a name="banner-icnn-96.ps.Z"</a><b><li>Revising Bayesian Network Parameters Using Backpropagation<br></b>Sowmya Ramachandran and Raymond J. Mooney<br><cite>Proceedings of the International Conference on NeuralNetworks (ICNN-96)</cite>, Special Session on Knowledge-Based ArtificialNeural Networks, Washington DC, June 1996. <p><blockquote>The problem of learning Bayesian networks with hidden variables is known tobe a hard problem. Even the simpler task of learning just the conditionalprobabilities on a Bayesian network with hidden variables is hard. In thispaper, we present an approach that learns the conditional probabilities ona Bayesian network with hidden variables by transforming it into amulti-layer feedforward neural network (ANN). The conditional probabilitiesare mapped onto weights in the ANN, which are then learned using standardbackpropagation techniques. To avoid the problem of exponentially largeANNs, we focus on Bayesian networks with noisy-or and noisy-andnodes. Experiments on real world classification problems demonstrate theeffectiveness of our technique.</blockquote><!WA8><a href="file://ftp.cs.utexas.edu/pub/mooney/papers/banner-icnn-96.ps.Z"><!WA9><img align=top src="http://www.cs.utexas.edu/users/ml/paper.xbm"></a><p><! ===========================================================================><a name="assert-dissertation-94.tar.Z" </a><b> <li> Automatic Student Modeling and Bug Library Construction using Theory Refinement <br> </b>  Paul T. Baffes <br>Ph.D. Thesis, Department of Computer Sciences, University of Texas atAustin, December, 1994.<p><blockquote>The history of computers in education can be characterized by acontinuing effort to construct intelligent tutorial programswhich can adapt to the individual needs of a student in aone-on-one setting. A critical component of these intelligenttutorials is a mechanism for modeling the conceptual state of thestudent so that the system is able to tailor its feedback to suitindividual strengths and weaknesses. The primary contribution ofthis research is a new student modeling technique which canautomatically capture novel student errors using only correctdomain knowledge, and can automatically compile trends acrossmultiple student models into bug libraries. This approach hasbeen implemented as a computer program, ASSERT, using a machinelearning technique called theory refinement which is a method forautomatically revising a knowledge base to be consistent with aset of examples. Using a knowledge base that correctly defines adomain and examples of a student's behavior in that domain,ASSERT models student errors by collecting any refinements to thecorrect knowledge base which are necessary to account for thestudent's behavior. The efficacy of the approach has beendemonstrated by evaluating ASSERT using 100 students tested on aclassification task using concepts from an introductory course onthe C++ programming language. Students who received feedback