R. S. Michalski, I. Mozetic, J. Hong, and N. Lavrac. The aq15 inductive learning system: An overview and experiments. In Proceedings of the American Association for Artificial intelligence Conference (AAAI), 1986.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....problems. Machine learning researchers would observe that Bongard problems are classi cation problems (another popular task is that of descriptive learning, for example discovering association rules [2, 1, 66] So, the range of possible propositional learning algorithms to consider includes AQ [47], TDIDT [57] like C4.5 [56] and CN2 [14, 13] Suppose we fancy the latter algorithm because it combines the advantages of AQ and TDIDT, i.e. it produces understandable rules, it is ecient and can cope with noisy data. So, we decide to base our rst order learner on CN2. Then we have also ....

.... the proposed methodology: Foil [58] RIBL [36] SRT [46] Tilde [9, 7] Warmr [26, 25] Maccent [24] jk CT learner [21] Claudien [20] Probabilistic Relational Models [45] Cohen s Flipper [16] 61] and RDBC [44] e.g. Quinlan s Foil can also be considered an upgrade of either Michalski s AQ [47] or CN2 [14, 13] RIBL upgrades the classical k nearest neighbor algorithm (using a rst order distance due to [6] SRT and Tilde upgrade the well known decision (and regression) tree paradigm incorporated in CART [12] and C4.5 [56, 57] Warmr upgrades Apriori [2, 1] Maccent upgrades the Maximum ....

R. Michalski, I. Mozetic, J. Hong, and N. Lavrac. The AQ15 inductive learning system: an overview and experiments. In Proceedings of IMAL 1986, Orsay, 1986. Universite de Paris-Sud.

....part or all of the rule evaluation [ Duda and Hart, 1973 ] We emphasize that the above four are only examples of a much larger space of useful generalized rule searches. 1. 1 RELATED WORK Rule learning and decision lists are a popular approach in machine learning, pioneered by [ Rivest, 1987, Michalski et al. 1986, Clark and Niblett, 1989 ] Generally, they search for the kind of conjunction ofliterals rules described above, concatenating them into chained if then elseif . statements. For classi cation, this paper gives very similar algorithms, except that we search a much wider space of possible rules ....

R. Michalski, I. Mozetic, J. Hong, and N. Lavrac. The aq15 inductive learning system: an overview and experiments. In In Proceedings of IMAL, 1986.

....that is forced to use the expert attributes only. 6 Conclusion Broadly termed constructive induction [17] the idea of augmenting an existing set of attributes is not new. Collins [3] describes a heuristic method of incorporating polynomial terms into a regression analysis. The AQ program [16] implements the capacity to introduce new attributes by specific combinations of existing ones. The idea is fundamental to the LINUS family of algorithms [15] where all possible attributes that can be constructed within a set of language restrictions are provided to a propositional learner. The ....

R. Michalski, I. Mozetic, J. Hong, and N. Lavrac. The AQ15 inductive learning system: an overview and experiments. In Proceedings of IMAL

....problems. Machine learning researchers would observe that Bongard problems are classification problems (another popular task is that of descriptive learning, for example discovering association rules [2, 1, 65] So, the range of possible propositional learning algorithms to consider includes AQ [46], TDIDT [56] like C4.5 [55] and CN2 [14, 13] Suppose we fancy the latter algorithm because it combines the advantages of AQ and TDIDT, i.e. it produces understandable rules, it is efficient and can cope with noisy data. So, we decide to base our first order learner on CN2. Then we have also ....

R. Michalski, I. Mozetic, J. Hong, and N. Lavrac. The AQ15 inductive learning system: an overview and experiments. In Proceedings of IMAL

....of these tools to real world databases requires efficient techniques to handle large data sets. Several techniques have been proposed: 1.0. 1 Sampling techniques where only a small subset of the database is used, such that e.g. neural net techniques [12] and traditional machine learning algorithms [13, 15] become tractable. A disadvantage of sampling is the inherent loss of information. 1.0.2 Incremental techniques which focus on discovering the hidden information with a single pass through the database [1] These techniques pre suppose a sizeable memory to maintain the learning or classification ....

Ryszard S. Michalski, Igor Mozetic, Jiarong Hong, and Nada Lavrac. The AQ15 inductive learning system: an overview and experiments. Technical Report UIUCDCS-R-86-1260, University of Illinois, July 1986.

....for the first case. Existing techniques for dealing with the first and third cases are both exclusively based on probability estimation. Among them, the Measure of Fit for dealing with the no match case and the Estimate of Probability for handling the multiple match case developed in AQ15 [16] have been widely adopted in knowledge discovery and data mining. 5 The Measure of Fit and Estimate of Probability methods perform quite well with problem domains where no real valued attributes are involved. However, when a problem contains attributes that take values from continuous domains ....

R. Michalski, I. Mozetic, J. Hong and N. Lavrac, The AQ15 Inductive Learning System: An Overview and Experiments, Proceedings of IMAL 1986, Universite de Paris-Sud, Orsay, 1986.

....datasets are chosen that may be benchmark datasets. As the actual subset is unknown in this case, the selected subset is tested for its accuracy with the help of any classifier suitable to the task. Typical classifiers used are: naive Bayesian classifier, Cart [7] ID3 [38] FRINGE [36] AQ15 [30], CN2 [11] and C4.5 [39] The achieved accuracy may be compared with that of some wellknown methods, and its efficacy is analyzed. In this section, we are trying to show an empirical comparison of representative feature selection methods. In our opinion, the second method of validation is not ....

Michalski, R.S., Mozetic, I., Hong, J. and Lavrac, N., The aq15 inductive learning system: An overview and experiments. Technical Report UIUCDCS-R-86-1260, University of Illinois, July 1986.

....it could have been written in a more declarative style that described pharmacophores directly. We see improvement in the form of the declarative bias as an area for further work with Progol in particular. For machine learning systems that take explicit background knowledge (e.g. metaDendral, AQ [20], ILP systems) it is widely accepted that a substantial knowledge engineering effort often is required to encode this knowledge. But we believe that a substantial knowledge engineering effort is required for the application of many machine learning systems that do not take explicit background ....

R. Michalski, I. Mozetic, J. Hong, and N. Lavrac. The AQ15 inductive learning system: an overview and experiments. In Proceedings of IMAL 1986, Orsay, 1986. Universit'e de Paris-Sud.

....Logic Programming (ILP) 38, 39] is one of the most general within the field of Machine Learning. ILP systems construct concept definitions (logic programs) from examples and a logical domain theory (background knowledge) This goes beyond the more established empirical learning framework [33, 49, 5, 6] because of the use of a first order relational logic together with background knowledge. It goes beyond the explanation based learning framework [34, 12] due to the lack of insistence on complete and correct background knowledge. The use of a relational logic formalism has allowed successful ....

R. Michalski, I. Mozetic, J. Hong, and N. Lavrac. The AQ15 inductive learning system: an overview and experiments. In Proceedings of IMAL 1986, Orsay, 1986. Universit'e de Paris-Sud.

....the latter s quality during search. Again in rule induction systems this technique is common, for example in the most recent ID3 descendants C4 [16] Assistant 86 [15] and by Niblett and Bratko [17] Niblett [18] gives a review of pre and post pruning techniques used for decision trees. AQ15 [19] employs a post pruning technique for production rules (termed rule truncation ) The use of post pruning of decision tree branches to generate production rules has been used by Corlett [20] and Quinlan [21] ffl Corroborative application of model components Another technique to prevent ....

....different weights attached to their decisions. This avoids heavy reliance on a specific, possibly unreliable, part of the system, allowing noisy effects to be over ridden and smoothed out by other model components. Statistical methods such as Bayesian techniques (e.g. 22] can be employed. AQ15 [19] performs weighted rule application in this way, and Quinlan [23] suggests how decision tree application can be made less brittle by introducing a degree of corroboration between decision tree branches. ffl The exception based paradigm For any system learning incrementally, noise presents the ....

R. S. Michalski, I. Mozetic, J. Hong, and N. Lavrac. The AQ15 inductive learning system : an overview and experiments. In Proceedings of IMAL 1986, Universit'e de Paris-Sud, Orsay, France, 1986.

....of foreign molecules. The final solution consists of the best B cell from each species. 4 The AQ Approach to Concept Learning We will compare the solutions produced by our coevolutionary immune system with those produced by AQ15, a symbolic inductive learning system developed by Michalski et al. [6]. This system is one of the latest in a series of AQ systems that constructs conjunctive descriptions from preclassified examples using an enhanced propositional calculus representation language. Each AQ concept description consists of a disjunction of conjunctive descriptions. Once a concept ....

....for each class of examples the system has been presented with, the system uses a conflict resolution procedure to discriminate between unclassified examples of one concept or another based on the strength of the match with the learned descriptions and the prior probability of the concepts. See [6] for more details concerning AQ conflict resolution and its method for constructing concept descriptions. 5 Experimental Study 5.1 Congressional Voting Records Data Set In this experimental study we evolve a political party classification system for members of the U.S. House of Representatives ....

R. S. Michalski, I. Mozetic, J. Hong, and N. Lavrac. The AQ15 inductive learning system: An overview and experiments. Technical Report UIUCDCS-R-86-1260, University of Illinois, Urbana-Champaign, IL, 1986.

....examples describing the behavior of this set of rules. Such behavioral examples are automatically derived from examples of the problem domain ; then inductive learning is used to extract new rules from these behavioral examples (many algorithms perform learning from examples ; see among others [9, 11, 8, 2, 15]) This paper is organized as follows: Section 2 formalizes the transition from examples about the problem domain into examples about the behavior of a given KB over the problem domain. This transition is performed by a redescription operator we call reduction. Section 3 shows how induction from ....

....the R Gamma s standing for the non matching of rule R. These descriptors enable completely new rules to be discovered, such as default rules: Default Class( GammaR Gamma 1 ( R Gamma L ( 2 Space limitations prohibit replicating pseudo code descriptions of induction; see [9, 2, 15] among others. 3.3 Requirements This induction based approach suffers the general machine learning requirements : it needs examples, typical examples and well described examples 3 . So we first need examples of conflicts among rules. Second, our set of descriptors must enable a sufficiently ....

[Article contains additional citation context not shown here]

Michalski R.S. Mozetic I. Hong J. Lavrac N. The AQ15 inductive learning system: an overview and experiment. In Proceedings of IMAL, 1986.

....building of oblique decision trees [HKS93, BFOS84] a predefined subset of functions of the initial attributes (e.g. linear combinations) is systematically explored during the building of the tree. 1. 2 Star algorithms The star algorithm repeatedly generalizes some selected examples termed seeds [Mic83, MMHL86]. The set of generalizations of a seed, termed star, consists of the best M rules that cover the seed and are optimal in the sense of some quality function ; both the number M of rules to retain and the optimality function, are supplied by the user. The user thereby takes in charge the control of ....

R.S. Michalski, I. Mozetic, J. Hong, and N. Lavrac. The AQ15 inductive learning system: an overview and experiment. In Proceedings of IMAL, 1986.

....5] as used for example in the systems c4 [6] and assistant [7] have proved to be effective methods of avoiding overfitting. The aq algorithm, however, is less easy to modify due to its dependence on specific training examples during its search. Existing implementations (e.g. aq11 [8] and aq15 [9]) deal with noisy data by using pre and post processing techniques while leaving the basic aq algorithm intact. Our objective in designing cn2 is to modify the aq algorithm itself in such a way that this dependence on specific examples is removed and the space of rules searched is increased. As a ....

....used for testing. The algorithms were all run on the same training data and their induced knowledge structures tested using the same test data. Five such tests were performed for each of the three domains, and the results were averaged. These data are thus identical to those used to test aq15 in [9], though the particular random 70 and 30 samples are different. Both cn2 and aqr were given a value of 15 for maxstar in all runs. 4.2.1 Three Medical Domains Table 4 summarizes the characteristics of the three medical domains used in the experiments. The first of these involved lymphography. ....

[Article contains additional citation context not shown here]

R. S. Michalski, I. Mozetic, J. Hong, and N. Lavrac. The AQ15 inductive learning system : an overview and experiments. In Proceedings of IMAL 1986, Orsay, France, 1986. Universit'e de Paris-Sud.

....For each concept, the 256 examples were randomly shuffled and then presented sequentially in batch incremental mode. This procedure was repeated 10 times (trials) for each concept and learning algorithm pair. For Domain 2, we used a well known natural database designed for diagnosing breast cancer (Michalski et al. 1986). This database has descriptions of cases for 286 patients, and each case (instance) is described in terms of 9 features. There is a small amount of noise in the database. Furthermore, the target concept is considerably more complex than any of the concepts in the nDmC world. For example, after ....

Michalski, R., I. Mozetic, J. Hong, and Lavrac, N., The AQ15 inductive learning system: An overview and experiments. University of Illinois Technical Report Number UIUCDCS-R-86-1260, 1986.

....[ Quinlan, 1986; Quinlan, 1993a ] is the best known and most succesful machine learning technique. It has been used to solve numerous practical problems. It employs a divide and conquer strategy, and in this it differs from its rule based competitors (such as CN2 [ Clark and Niblett, 1989 ] AQ [ Michalski et al. 1986 ] which are based on covering strategies (cf. Bostrom, 1995 ] Within attribute value learning (or propositional concept learning) TDIDT is more popular than the covering approach. Yet, within first order approaches to concept learning, only a few learning systems have made use of decision ....

R. Michalski, I. Mozetic, J. Hong, and N. Lavrac. The AQ15 inductive learning system: an overview and experiments. In Proceedings of IMAL 1986, Orsay, 1986. Universit'e de Paris-Sud.

....for category prediction by the machine learning community. They use heuristic methods to find simple rule sets to explain training data well, and are generally categorized into two groups [17] simultaneously covering algorithms namely C4.5 [18] and sequential covering algorithms such as AQ15 [15] and CN2 [6] Most algorithms generate simple and accurate rule sets that cover all training data, but these rule sets may not be robust in the presence of missing values as we will discuss in this paper. There are many proposals for improving predictive accuracy of traditional classifiers, among ....

R. Michalski, I. Mozetic, J. Hong, and N. Lavrac. The AQ15 inductive learning system: an overview and experiments. In Proceedings of IMAL 1986.

....prediction of object sequences (SPARC) and derivation of equations and rules characterizing data about physical processes (ABACUS) Each of these programs is directly applicable to conceptual data exploration. For example, the rules in Figure 2. 1 were generated by the AQ15 rule module [MMHL86], HMM86] from a set of positive and negative examples of Class 1 of robot figures. AQ15 learns attributional descriptions of entities, i.e. descriptions involving only their attributes. More general descriptions, structural or relational. also involve relationships among components of the ....

....of constructive induction (e.g. WM94] Each attribute is assigned a domain and a type. The domain specifies the set of all legal values that the attribute can be assigned in the table. The type defines the ordering (if any) of the values in the domain. For example, the AQ15 learning program [MMHL86] allows four types of attributes: nominal (no order) linear (total order) cyclic (cyclic total order) and structured (hierarchical order; see [KM96] The attribute type determines the kinds of operations that are allowed on this attribute s values during a learning process. Entries in each row ....

Michalski, R. S., Mozetic, I, Hong, J. and Lavrac, N. The AQ15 Inductive Learning System: An Overview and Experiments. ISG Report 86-20, UIUCDCS-R-86-1260, Department of Computer Science, University of Illinois, Urbana, 1986.

....the induced rules, whereby their interpretation involves the use of weights and probabilities instead of solely boolean values, thus exploiting to the maximum information contained in the training data. This technique also used by AQ11 and is on of the important features of the AQ15 algorithm, see [9], 10] Misclassification by an erroneous rule may be overridden by other rules, whose conditions are nearly met and which have higher weight attached. 2.2.3. Rule Truncation A third method is, after induction, to remove the rules which represent the weakest correlations found between attributes ....

....knowledge can be used to reduce problems of description language and noise. Explanation based generalization, for example [16] 17] constrains generalization to be performed only where the generalization s validity can be proved. DISCIPLE [18] uses explanations to guide generalization. AQ15 [9] allows the user to provide background knowledge to assist in induction. This paper presents a description and empirical evaluation of a new induction system based on the 4th of these techniques, involving the relaxing of the requirement of complete consistency of rules with the training data ....

[Article contains additional citation context not shown here]

Michalski R., Mozetic I., Hong J., Lavrac N. (1986a) The AQ15 inductive learning system : an overview and experiments Proceedings of IMAL 1986, Orsay: Universite de Paris-Sud.

No context found.

R. S. Michalski, I. Mozetic, J. Hong, and N. Lavrac. The aq15 inductive learning system: An overview and experiments. In Proceedings of the American Association for Artificial intelligence Conference (AAAI), 1986.

No context found.

Michalski, R., Mozetic, I., Hong, J., Lavrac, N.: The AQ15 Inductive Learning System: an Overview and Experiments, in Proceedings of International Meeting on Advances in Learning, 1986

No context found.

R.S. Michalski, I. Mozetic, J. Hong, and N. Lavrac. The AQ15 inductive learning system: an overview and experiment. In Proceedings of IMAL, 1986.

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