Mooney, R., Shavlik, J., Towell, G., & Gove, A. (1989). An experimental comparison of symbolic and connectionist learning algorithms. Proceedings of the Eleventh International Joint Conference on Artificial Intelligence (pp. 775-780). Detroit, Michigan: Morgan Kaufmann.  Home/Search   Document Not in Database   Summary   Related Articles   Check

.... in both fields as evidenced by the sizes of the two leading conferences (IJCNN and IJCAI) And relatively recently, researchers from both sides of the fence have begun to carry out comparative work to try to establish the precise strengths and weaknesses of symbolic and subsymbolic mechanisms (e.g. [7, 8, 9]) Initial results seem to suggest that in the field of learning, symbolic and subsymbolic algorithms are of comparable performance overall. Of course the existence of such comparative studies is not enough to prevent extremists in one paradigm or the other making exaggerated claims about the ....

Mooney, R., Shavlik, J., Towell, G. and Gove, A. (1989). An experimental comparison of symbolic and connectionist learning algorithms. Proceedings of the Eleventh International Joint Conference On Artificial Intelligence (pp. 775-780). Morgan Kaufmann.

....The comparison was based on the performance accuracy of descriptions as applied to testing examples and the training time. Their conclusion was that the neural net gave a better performance, but required a significantly longer mining time and more training examples than ID3. Mooney et al. [4] compared ID3 with perceptnon and backpropagation algorithms using the domain of soybean diseases, chess end games, audiological disorders, and the Nettalk data set. Their conclusion wfis that the accuracy of classifying new examples was about the same for all the three systems, but the neural net ....

Mooney, R.J., Shavlik, J., Towell, G., and Gove, A. 1989. An Experimental Comparison of Symbolic and Connectionist Learning Algorithms. In Proceedings of HCAI-89. Detroit, MI: Morgan Kaufmann.

....because a training set learner learns on its own in that it is responsible for extracting from the training set the critical features of A without human intervention. These critical features are used in turn to approximate A. Empirical studies have shown that good results can be achieved with TSL [Mooney et al. 90, Kibler and Aha 87, Sejnowski and Rosenberg 87] However, there are several problems with TSL. Training set learners (e.g. backpropagation) are typically slow as they may require several passes over the training set. Also, there is no guarantee that, given any arbitrary training set, the system ....

....are different, AI ML and NN have the common goal of designing systems capable of learning certain human tasks. It has thus become customary to merge 4 the two disciplines into the area of Machine Learning. In blending the two approaches, the community opened the door to comparative studies [Mooney et al. 89] as well as to attempts at designing systems combining them [Shavlik and Towell 89, Hall and Romaniuk 90, Barker 93, This Thesis] In the mean time, the recently renewed interest in machine learning has given rise to a new taxonomy of learning mechanisms in general. Remy 88, Carbonell et al. ....

[Article contains additional citation context not shown here]

Mooney, R., Shavlik, J., Towell, G., and Gove, A. An Experimental Comparison of Symbolic and Connectionist Learning Algorithms. In Proceedings of the Eleventh International Joint Conference on Artificial Intelligence,

....this process itself may eventually become automated. Given an application A, inductive learners typically extract the critical features of A from the training set, without human intervention. Empirical studies show that good results can be achieved with such inductive systems (see, for example [KIB87, MOO89, QUI86, SEJ87, ZAR94]) However useful, IL has several important drawbacks. Many such systems (e.g. backpropagation) are typically slow as they may require many passes over the training set to achieve a reasonable level of performance. Also, there is no guarantee that given an arbitrary training set, the system will ....

....set for A. It is also philosophically attractive because a training set learner extracts the critical features of A from the training set, without human intervention. These critical features are used in turn to approximate A. Empirical studies show that good results can be achieved with TSL [MOO89, SEJ87]. However, TSL has several drawbacks. Training set learners (e.g. backpropagation) are typically slow as they may require many passes over the training set. Also, there is no guarantee that, given an arbitrary training set, the system will find enough good critical features to get a reasonable ....

Mooney, R., Shavlik, J., Towell, G., and Gove, A. (1989). An Experimental Comparison of Symbolic and Connectionist Learning Algorithms. In Proceedings of the 11th International Joint Conference on Artificial Intelligence, 775-780.

.... networks [ Huang et al. 1991, Fahlman, 1991, Xu et al. 1991, Ersoy and Hong, 1991 ] There have been fewer, 3 but still many, inter subject studies involving algorithms from two or more of these fields, e.g. Weiss and Kapouleas, 1989, Weiss and Kulikowski, 1991 ] Ripley, 1992 ] Mooney et al. 1989, Shavlik et al. 1991 ] Fisher and McKusick, 1989 ] Thrun et al. 1991 ] Kirkwood et al. 1989 ] Tsaptsinos et al. 1990 ] Spikovska and Reid, 1990 ] Atlas et al. 1991 ] and [ Gorman and Sejnowski, 1988 ] The comparisons by [ Weiss and Kapouleas, 1989, Weiss and ....

.... (PVM [ Weiss et al. 1990 ] and CART [ Breiman et al. 1984 ] were more accurate (on most datasets) than back propagation or a number of statistical algorithms (e.g. linear and logistic discriminants) In similar trials involving ID3 [ Quinlan, 1986 ] a perceptron, and back propagation [ Mooney et al. 1989, Shavlik et al. 1991 ] and [ Fisher and McKusick, 1989 ] reported that back propagation was at least as accurate as ID3 and the perceptron algorithm. Ripley [1992] compared a diverse set of algorithms on two related problems of Tetse fly distribution, and found that the nearest neighbor and ....

[Article contains additional citation context not shown here]

R. Mooney, J. Shavlik, G. Towell, and A. Gove. An experimental comparison of symbolic and connectionist learning algorithms (vol 1). IJCAI 89: proceedings of the eleventh international joint conference on artificial intelligence. 20-25 August 1989, Detroit, MI., pp 775 -- 780, San Mateo, CA. Morgan Kaufmann. 54

....which is true (1) if and only if the attribute has taken on the particular value in the example, and false (0) otherwise (Hampson Volper, 1986) This scheme makes it possible to describe examples by a mix of both boolean and symbolic attributes. In fact, in the context of decision trees, Mooney et al. (1989) have observed that mapping many valued attributes to two valued propositional attributes consistently increases classification accuracy. 14 5.3 Inconsistent Examples Two examples are inconsistent if they are described by the same attribute values but have different class labels. When ....

Mooney, R., Shavlik, J., Towell, G., & Gove, A. (1989). An experimental comparison of symbolic and connectionist learning algorithms. Proceedings of the Eleventh International Joint Conference on Artificial Intelligence (pp. 775-780). Detroit, Michigan: Morgan Kaufmann.

....that they are relatively easy for humans to understand. Actually, they have been used by human experts to express and process their knowledge in a wide variety of domains. In addition, compared with other theory description languages, they perform reasonably well in many domains [Quinlan, 1988b; Mooney, Shavlik, Towell, and Gove, 1989; Michie, Spiegelhalter, and Taylor, 1994] A decision tree can be recursively defined as either a leaf (terminal node) that names a class, or a decision node (non terminal node or internal node) that specifies an attributebased test with a branch to another decision tree (called a subtree) for ....

R. Mooney, J. Shavlik, G. Towell, and A. Gove, An experimental comparison of symbolic and connectionist learning algorithms. Proceedings of the Eleventh International Joint Conference on Artificial Intelligence, San Mateo, CA: Morgan Kaufmann, 775-780.

....for that block. Choosing a test that immediately partitions a set of examples into more than two blocks is more aggressive. By partitioning more conservatively, one keeps a larger number of examples available in each block, which is important if additional partitioning will be done in that block. Mooney, Shavlik, Towell, and Gove (1989) found that recoding discrete variables as propositional variables improved classification accuracy for the ID3 decision tree induction algorithm. Breiman, Friedman, Olshen, and Stone (1984) employ binary tests in the CART decision tree induction program. Fayyad (1991) observes that for numeric ....

Mooney, R., Shavlik, J., Towell, G., & Gove, A. (1989). An experimental comparison of symbolic and connectionist learning algorithms. Proceedings of the Eleventh International Joint Conference on Artificial Intelligence (pp. 775-780). Detroit, Michigan: Morgan Kaufmann.

....where the suffix y affects the pronunciation of the first vowel (among other things) Due to its difficulty, there has been a lot of interest in this specific task as a test bench for machine learning algorithms, both connectionist and not. Examples of this kind of work include [4] 3] [12], 13] and [14] Most of this work tends to ignore any additional knowledge about the problem domain but the examples. This is, of course, a good viewpoint when one is demonstrating the validity of a machine learning method, or when no such knowledge does exist. We, however, introduce a small ....

R. Mooney, J. Shavlik, G. Towell, and A. Gove. An experimental comparison of symbolic and connectionist learning algorithms. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI89), pages 775--780, Detroit, Michigan, USA, August 20-25 1989.

....we feel there are distinct advantages to including neural networks in our study. First, previous empirical studies have demonstrated that individual neural networks produce highly accurate classifiers that are sometimes more accurate than corresponding decision trees (Fisher McKusick, 1989; Mooney, Shavlik, Towell, Gove, 1989). Second, neural networks have been extensively applied across numerous domains (Arbib, 1995) Finally, by studying neural networks in addition to decision trees we can examine how Bagging and Boosting are influenced by the learning algorithm, giving further insight into the general ....

Mooney, R., Shavlik, J., Towell, G., & Gove, A. (1989). An experimental comparison of symbolic and connectionist learning algorithms. In Proceedings of the Eleventh International Joint Conference on Artificial Intelligence, pp. 775--780 Detroit, MI.

....for an incremental approach to place the hyperplanes optimally as opposed to the search methods used in standard decision trees. 1 To appear in the Proceedings of IJCNN, Seattle, 1991 Recently there have been many empirical studies comparing the performance of Neural Networks and Decision Trees [10, 11, 12, 13]. One particularly interesting test bed for pattern classifiers is Speech Recognition. The relative performances of MLPs and Decision Trees have been evaluated on a speaker independent vowel recognition task in [11, 12] These studies show that while MLPs sometimes have a smaller classification ....

R. Mooney, J. Shavlik, G. Towell, and A. Gove, "An Experimental Comparison of Symbolic and Connectionist Learning Algorithms," in Proceedings of the Eleventh International Joint Conference on Artificial Intelligence, 1989.

....machine learning methods typically are justified by noting that they can capture com SCALING UP INDUCTIVE ALGORITHMS 7 plex, non linear relationships from data. Nevertheless, research on both symbolic and neural learning has shown that simple models perform well on many problems. For example, Shavlik et al. 1991) show that with certain qualifications, the accuracy of the perceptron is hardly distinguishable from the more complicated learning algorithms. One level decision trees, also known as decision stumps, are simple mappings from the values of one attribute to class labels. Decision stumps also have ....

Shavlik, J. W., R. J. Mooney, and G. G. Towell (1991). An experimental comparison of symbolic and connectionist learning algorithms. Machine Learning 6(2), 111--143.

....approach to data driven knowledge discovery from labeled examples. Pattern classifiers that are induced by rule learning algorithms are often simpler and easier to comprehend by humans than those induced using genetic programming or most neural network approaches. Yet, results of experiments [Mooney et al. 1989] indicate that the classification accuracies of the classifiers induced using different approaches are often comparable. A variety of algorithms for rule induction from labeled examples have been proposed in the literature. Some of them first construct a decision tree e.g. using C4.5 [Quinlan, ....

Mooney, R., Shavlik, J., Towell, G., and Gove, A. (1989). An experimental comparison of symbolic and connectionist learning algorithms. In Proceedings of the Eleventh International Joint Conference on Artificial Intelligence, pages 775--780. Morgan Kauffman.

....are drawn illustrating the functioning of the described approaches. The task is also in focus in Section 3. Grapheme phoneme conversion is a well known benchmark task in machine learning (Sejnowski and Rosenberg, 1987; Stanfill and Waltz, 1986; Stanfill, 1987; Lehnert, 1987; Wolpert, 1989; Shavlik, Mooney, and Towell, 1991; Dietterich, Hild, and Bakiri, 1995) We define the task as the conversion of fixed sized instances representing parts of words to a class representing the phoneme of the instance s middle letter. To generate the instances, windowing is used (Sejnowski and Rosenberg, 1987) Table 1 displays four ....

Shavlik, J. W., R. J. Mooney, and G. G. Towell. 1991. An experimental comparison of symbolic and connectionist learning algorithms. Machine Learning, 6:111--143.

.... grapheme phoneme conversion with stress assignment Converting written words to stressed phonemic transcription, i.e. word pronunciation, is a well known benchmark task in machine learning (Sejnowski and Rosenberg, 1987; Stanfill and Waltz, 1986; Stanfill, 1987; Lehnert, 1987; Wolpert, 1989; 8 Shavlik, Mooney, and Towell, 1991; Dietterich, Hild, and Bakiri, 1995) We define the task as the conversion of fixed sized instances representing parts of words to a class representing the phoneme and the stress marker of the instance s middle letter. We henceforth refer to the task as gs, an acronym of grapheme phoneme ....

Shavlik, J. W., R. J. Mooney, and G. G. Towell. 1991. An experimental comparison of symbolic and connectionist learning algorithms. Machine Learning, 6:111--143.

....for classifier induction because they promise highly parallel (and hence very fast) realtime classification. They also provide a novel representational formalism, which leads to improved performance on some tasks, as demonstrated by several recent empirical studies [ Fisher and McKusick, 1989, Mooney et al. 1989, Atlas et al. 1990c, Atlas et al. 1990d, Cole et al. 1990, Atlas et al. 1990a, Dietterich et al. 1990, Fahlman and Lebiere, 1990b, Pratt and Norton, 1990, Weiss and Kulikowski, 1991, Shavlik et al. 1991, Thrun et al. 1991 ] However, these same studies indicate that neural network ....

Raymond J. Mooney, J. W. Shavlik, G. G. Towell, and A. Gove. An experimental comparison of symbolic and connectionist learning algorithms. In Proceedings of the Eleventh International Joint Conference on Artificial Intelligence, pages 775--780, August 1989.

.... shown that neural networks often produce competitive, and sometimes superior, results ( Weiss and Kulikowski, 1991, Shavlik et al. 1991, Thrun et al. 1991, Atlas et al. 1990b, Atlas et al. 1990c, Cole et al. 1990, Atlas et al. 1990a, Dietterich et al. 1990, Fisher and McKusick, 1989, Mooney et al. 1989, Pratt, 1990 ] Neural network training techniques still have room for improvement, however. Though they eventually achieve good performance levels, neural networks often require more computing time than competing methods (cf. Maren et al. 1990, Page 92 ] Waibel et al. 1989 ] Hertz ....

Raymond J. Mooney, J. W. Shavlik, G. G. Towell, and A. Gove. An experimental comparison of symbolic and connectionist learning algorithms. In Proceedings of the Eleventh International Joint Conference on Artificial Intelligence, pages 775--780, August 1989.

....Yes No Duration = 60 Duration = 30 Duration = 90 Figure 5. Part of an ID3 Decision Tree for Predicting Meeting Duration. propagation (Rumelhart et al. 1986) Both these methods have previously been shown to be effective in learning from examples (see, for example, Fisher and McKusick, 1989) (Mooney et al. 1989)) ID3 ID3 (Quinlan, 1986) Quinlan, 1987) is an inductive algorithm which builds a decision tree for classification of objects. Given a set of objects with known classes and described by a fixed set of features, a decision tree is produced which predicts the class of a new object described by ....

....by ID3 for Location. of epochs. Similarly, there are many variations of ID3 which may cope better with noisy data, overfitting of data, etc. Thus we cannot draw too much from the comparison of these learning methods. However, our initial results confirm those of (Fisher and McKusick, 1989) (Mooney et al. 1989): Backpropagation performs slightly better than ID3 on average with noisy data, but takes much longer to train. We are currently investigating the possibility of initializing the network from the hand coded rules or ID3 generated rules, to try to improve the speed and accuracy of ....

Mooney, R., Shavlik, J., Towell, G., and Gove, A. (1989). An experimental comparison of symbolic and connectionist learning algorithms.

....numbers of processing units, each with minimal capability, operate without global control and together produce intelligent behavior. ANNs have proven successful on low level tasks, such as perception [McClelland81] and signal processing [Ahalt89] as well as on classification and diagnostic tasks [Mooney89b]. Each approach has its strengths and weaknesses. An approach that combines the two styles, building on each s strengths and overcoming each s weaknesses, is described and a successful implementation reported. The approach is not intended to be plausible at the level of neurophysiology. Rather, it ....

....belonging to a particular class, is provided to the ANN. Its task is to determine a procedure for classifying future examples. Impressive performance has been achieved on problems such as converting text to speech [Sejnowski87] evaluating moves in backgammon [Tesauro89] and diagnosing diseases [Mooney89b]. Several algorithms have been developed for training an ANN to classify new examples (e.g. Hinton86, Rosenblatt62, Rumelhart86b] Training is accomplished by analyzing a set of provided, pre classified examples to determine how to appropriately modify the connection weights so that these ....

[Article contains additional citation context not shown here]

R. J. Mooney, J. W. Shavlik, G. Towell and A. Gove, "An Experimental Comparison of Symbolic and Connectionist Learning Algorithms," Proceedings of the Eleventh International Joint Conference on Artificial Intelligence, Detroit, MI, August 1989, pp. 775-780.

No context found.

Mooney, R., Shavlik, J., Towell, G., & Gove, A. (1989). An experimental comparison of symbolic and connectionist learning algorithms. Proceedings of the Eleventh International Joint Conference on Artificial Intelligence (pp. 775-780). Detroit, Michigan: Morgan Kaufmann.

No context found.

R. Mooney, J. Shavlik, G. Towell, and A. Gove. An experimental comparison of symbolic and connectionist learning algorithms. In IJCAI-89 Proceedings of the Eleventh International Joint Conference on Artificial Intelligence, volume 1, pages 775--780, 1989.

No context found.

R. Mooney, J. Shavlik, G. Towell, and A. Gove. An experimental comparison of symbolic and connectionist learning algorithms. In IJCAI-89, pages 775--780, 1989.

No context found.

Mooney R., Shavlik J., et al. 1989. An experimental comparison of symbolic and connectionist learning algorithms. In Proceedings of the 12 International Joint Conference on Artificial Intelligence (IJCAI-89), 775-780.

No context found.

R. Mooney, J. Shavlik, G. Towell, and A. Gove. An experimental comparison of symbolic and connectionist learning algorithms. In International Joint Conference on Artificial Intelligence, 1989.

No context found.

. An Experimental Comparison of Symbolic and Connectionist Learning Algorithms. Machine Learning,

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