J.J. Mahoney, `Combining Symbolic and Connectionist Learning Methods to Refine Certainty-Factor Rule-Bases', PhD Dissertation, University of Texas at Austin, 1996.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....it as possible. 1 INTRODUCTION There has been extensive research activity at combining (or integrating) the symbolic and the connectionist approaches for knowledge representation in expert systems [7, 8, 10] Especially, there are a number of efforts combining symbolic rules and neural networks [2, 3, 6, 9]. They give pre eminence to connectionism and use a neural network as a knowledge base. The main objective is to reduce knowledge elicitation from experts to a minimum. In such approaches, connectionism is mainly used as a means for refining an initial background rule base. Integration with ....

J.J. Mahoney, `Combining Symbolic and Connectionist Learning Methods to Refine Certainty-Factor Rule-Bases', PhD Dissertation, University of Texas at Austin, 1996.

....it in a more profitable manner. This can be done by using the theory as a resource for constructive induction (Pazzani Kibler, 1992; Donoho Rendell, 1995; Ortega Fisher, 1995; Koppel Engelson, 1996) or by numerical refinement of probabilistic theories 1 (Mahoney Mooney, 1994; Mahoney, 1996; Buntine, 1991; Lam Bacchus, 1994; Russell, Binder, Koller, Kanazawa, 1995; Ramachandran Mooney, 1998) or, most relevant to this paper, by interpreting a logical theory in a probabilistic manner (Towell Shavlik, 1993; Koppel, Feldman, Segre, 1994b; Ortega, 1995) All of these methods ....

Mahoney, J. J. (1996). Combining Symbolic and Connectionist Learning to Revise Certainty-Factor Rule Bases. Ph.D. thesis, Department of Computer Sciences, University of Texas, Austin, TX. Also appears as Artificial Intelligence Laboratory Technical Report AI 96-260.

....represent expert theories that have been previously written in a simplified form of Horn clause logic. There are a number of systems that use neural network encodings to refine a variety of representation types, including finite state automata (Maclin Shavlik, 1993) certainty factor rule bases (Mahoney, 1996), and first order horn clause logic (Towell et al., 1990) indigent uses the Kbann (Towell et al., 1990) method for encoding domain theories. There has also been work done on genetically 3 refining neural network topologies. Opitz (1995) describes the Regent system, an extension to Kbann, which ....

....an earlier extension of Kbann. It does modify the set of input features. The system is significantly different from indigent, however, since it chooses the input nodes to add through induction on erroneous hidden nodes and the feature set, rather than utilizing genetic search. The Rapture system (Mahoney, 1996) was developed to genetically refine certainty factor rule bases. Like Regent, Rapture genetically refines the topology of hidden nodes in the networks. Rapture does attempt to improve the set of input features by adding relevant (as judged by the information gain metric (Quinlan, 1986) nodes if ....

[Article contains additional citation context not shown here]

Mahoney, J. (1996). Combining Symbolic and Connectionist Learning Methods to Refine CertainttyFactor Rule-Bases. PhD thesis, University of Texas, Austin.

....as neural networks can be refined through the addition or removal of rules in the knowledge base. A number of different representations of human knowledge have been encoded in the form of neural networks, including deterministic finite automata (DFAs) OG96, MS93] certainty factor rule bases [Mah96], pushdown automata [DGS92] and modified first order logic rule bases [TSN90] Most of these systems translate the domain knowledge into a network topology and then refine the network s biases and weights through standard neural network techniques in order to revise the theory. The revision ....

....enhance diversity as well as accuracy. The output of Addemup is the weighted average of the outputs of the individual networks in its ensemble. While Addemup fails to develop a single best network, its ensembles show improvements in accuracy over Regent and some improvement over indigent. Rapture [Mah96] was designed to implement certainty or probability factor rule bases as neural networks. Rapture is also able to refine the topologies of these networks. However, the method that was devised for modifying the topology is significantly different from either tnt indigent or Regent. Rapture ....

J. Mahoney. Combining Symbolic and Connectionist Learning Methods to Refine Certaintty-Factor Rule-Bases. PhD thesis, University of Texas, Austin, 1996.

....represent expert theories that have been previously written in a simplified form of Horn clause logic. There are a number of systems that use neural network encodings to refine a variety of representation types, including finite state automata (Maclin Shavlik 1993) certainty factor rule bases (Mahoney 1996), and firstorder horn clause logic (Towell, Shavlik, Noordewier 1990) indigent uses the Kbann (Towell, Shavlik, Noordewier 1990) method for encoding domain theories. There has also been work done on genetically refining neural network topologies. The Regent (Opitz 1995) system is an extension ....

....an earlier extension of Kbann . It does modify the set of input features. The system is significantly different from indigent, however, since it chooses the input nodes to add through induction on erroneous hidden nodes and the feature set, rather than utilizing genetic search. The Rapture system (Mahoney 1996) was developed to genetically refine certainty factor rule bases. Like Regent, Rapture genetically refines the topology of hidden nodes in the networks. Rapture does attempt to improve the set of input features by adding relevant (as judged by Quinlan s (Quinlan 1986) information gain metric) ....

Mahoney, J. 1996. Combining Symbolic and Connectionist Learning Methods to Refine Certaintty-Factor Rule-Bases. Ph.D. Dissertation, University of Texas, Austin.

....the less restricting representation of neural networks (Donoho Rendell, 1995) Also, Regent is able to further reconfigure the structure of the domain with genetic algorithms. Many authors have reported results using varying subsets of the splice junction domain (e.g. Donoho and Rendell 1995; Mahoney 1996; Neri and Saitta 1996, and Towell and Shavlik 1994) While these authors used different training set sizes, it is nevertheless worthwhile to qualitatively discuss some of their conclusions here. Towell and Shavlik (1994) compared Kbann with numerous machine learning algorithms where each ....

....examples; Kbann s generalization ability compared favorably with these algorithms on the splice domain and Regent, in turn, compared favorably with Kbann in this article. Donoho and Rendell (1995) showed their purely symbolic approach converged to the performance of Kbann at around 200 examples. Mahoney (1996) showed, using training set sizes of up to 400 examples, that his Rapture algorithm generalized better than Kbann on this domain; his results look similar to those of Regent. Finally, Neri and Saitta (1996) showed that the generalization ability of the GA based Regal compares favorably to other ....

Mahoney, J. (1996). Combining Symbolic and Connectionist Learning Methods to Refine Certainty-Factor Rule-Bases. Ph.D. thesis, University of Texas, Austin, TX.

No context found.

J. Mahoney. Combining Symbolic and Connectionist Learning Methods to Refine Certaintty-Factor Rule-Bases. PhD thesis, University of Texas, Austin, 1996.

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