M. Botta and A. Giordana. SMART+: A multi-strategy learning tool. In IJCAI93, Proceedings of the Thirteenth International Joint Conference on Artificial Intelligence, volume 2, Chambery, France, 1993.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....sum function can be used for the output neuron. 9.0.0.13 Concluding Remarks. The symbolic interpretation of an RBFN allows a wide range of symbolic learning algorithms to be applied in order to initialize the basis functions. Decision trees [17] or symbolic induction systems such as SMART [15] can be used to construct the layout from a sample set of data. Alternatively, if domain knowledge is available, e.g. from an expert, it can be directly inserted into the network. Gradient descent provides a technique for refining knowledge with data. Finally, it is possible to exploit symbolic ....

M. Botta and A. Giordana. SMART+: A multi-strategy learning tool. In IJCAI93, Proceedings of the Thirteenth International Joint Conference on Artificial Intelligence, volume 2, Chambery, France, 1993.

....that may appear in more than one rule. Concluding Remarks The symbolic interpretation of an RBFN allows a wide range of symbolic learning algorithms to be applied in order to initialize the basis functions. Decision trees [Breiman et al. 1984] or symbolic induction systems such as SMART 32 [Botta and Giordana, 1993] can be used to construct the layout from a sample set of data. Alternatively, if domain knowledge is available, e.g. from an expert, it can be directly inserted into the network. Gradient descent provides a technique for refining knowledge with data. Finally, it is possible to exploit symbolic ....

....by DCL and DRT and the ones generated by CART and k Means. Finally, Table 5. 5 reports a comparison with some of the best results reported in the literature, such as ANFIS [Jang, 1993] a fuzzy neural network de 85 signed by Jang and an F RBFN generated using a symbolic learner called SMART [Botta and Giordana, 1993] which allows a domain theory to be take into account. However, considering that on line learning is inherently suboptimal with respect to off line learning, the results of DCL and DRT appear satisfactory in comparison to the ones reported in the literature. Table 5.3: Performance after training ....

Botta, M. and Giordana, A. (1993). SMART+: A multistrategy learning tool. In IJCAI-93, Proceedings of the Thirteenth International Joint Conference on Artificial Intelligence, volume 2, Chamb'ery, France.

....probably greatest arises from the nature of the elementary operations. Instead of a sequence of discrete actions, the control involved in the supervision of an elementary operation (e.g. a force controlled movement, contour tracking, etc. requires the learning of a continuous numerical function [1, 3, 11]. Therefore, completeness or sufficiency of the examples is much more important, but much more difficult to achieve than on the task level. Even bigger problems arise from the nonoptimal input provided by the user. While this can relatively easily be detected by the user himself on the task level, ....

....this area exist. They range from the initial generation of a neural network or a fuzzy controller on the base of some existing controller and later on line tuning on the field over the interpretation of the whole control task as a markovian model to the direct playback of a compliance trajectory [3, 11, 22, 23]. In each of these approaches a lot of simplifying assumptions that do not hold in reality are made. A possible way to overcome these problems might be a better integration of the user into the acquistion system, not only by allowing him to supervise the whole process, but especially by providing ....

M. Botta, A. Giordana. SMART+: A multistrategy learning tool. In: Proc. of the Int. Joint Conference on Artificial Intelligence (IJCAI '33], Chambery, France, 1993

.... In opposition, FOIL, FOCL and PROGOL, among others, aim at finding the best hypothesis covering a training example, according to the more or less greedy optimization of a numerical criterion (quantity of information for FOIL and FOCL, MDL principle for PROGOL) To a lesser extent, ML Smart [1, 3] and REGAL [7] also look for concise theories. Let us focus now on building the set D(Ex;Ce) of hypotheses generalizing Ex and rejecting Ce, depending on the hypothesis language. 2.3 Attribute value learning In an attribute value language, the construction of D(Ex;Ce) is straightforward. ....

....in Sigma Ex;Ce would minimize the sum of the distances between atom i in Ex and atom oe(i) in Ce. As noted in [33] the description of an atom can be handled as a single treestructured feature since the element of an atom commands its atom type (e.g. the atom type of a hydrogen atom is in [1,3] whereas the atom type of a carbon atom is in [21,24] and the atom type similarly commands its electric charge. Defining a distance between any two atoms thus is straightforward: the distance of two atoms having same atom type is the difference of their electric charges; otherwise, if the atoms ....

M. Botta and A. Giordana. Smart+ : A multi-strategy learning tool. In Proceedings of IJCAI-93, pages 937--943. Morgan Kaufmann, 1993.

....for each linguistic variable h i 2 H (our target concepts) So the problem of learning a continuous control function will be mapped onto the problem of learning #H classification theories. In the experiment described here, SMART , an inductive system capable of inducing first order logic relations [5], has been used to accomplish the inductive task. Induction follows a general to specific strategy similar to the one used by other systems (e.g. FOIL [16] FOCL [15] and the search can be biased using a domain theory. As a particular case, a partial concept description can be suggested. This ....

....1.0) DF y ) Dfz(k: 9, 10, 0.5] fuzzy( k, 1.0) DF z ) Dmx(k: 400, 400, 10] fuzzy( k, 12) DM x ) Dmy(k: 400, 400, 10] fuzzy( k, 12) DM y ) Dmz(k: 400, 400, 10] fuzzy( k, 12) DM z ) Ifx(k: 10, 10, 1] fuzzy(k, 1.1) F x ) Ify(k: 10, 10, 1] fuzzy(k, 1. 1) F y ) Ifz(k:[ 61, 25, 5]) fuzzy(k, 6.0) F z ) Imx(k: 3000, 500, 50] fuzzy(k, 70) M x ) Imy(k: 2000, 3000, 50] fuzzy(k, 70) M x ) Imz(k: 180, 120, 10] fuzzy(k, 11) M x ) Table 2: linguistic variables defined for the input signals. the recorded traces, are reported. Table 3 also makes a comparison ....

M. Botta and A. Giordana. SMART+: a multi-strategy learning tool. In Proc. of the International Joint Conference on Artificial Intelligence, IJCAI-93, Chambery, France, 1993.

.... tractable (Muggleton De Raedt 1994) Top down learners are driven by optimality criterions (e.g. quantity of information, Gini criterion, MDL principle) allowing one both to cope with noisy data and to restrict the search to optimal or near optimal regions of the hypothesis space (Quinlan 1993, Botta Giordana 1993). Divide andconquer algorithms also employ learning biases to deal And Equipe I A, LRI, Universit e Paris Sud, 91405 Orsay, FRANCE. with noise, through specifying the maximal number of inconsistencies acceptable for a hypothesis (Michalski 1983, Ganascia 1993, Muggleton 1995) or the noise ....

Botta, M. and A. Giordana. Smart+ : A multistrategy learning tool. In Proceedings of IJCAI-93, pages 937--943. Morgan Kaufmann, 1993.

....on finite domains. Its application to continuous informations involves a discretization step, i.e. a change of representation of the initial problem. This change has a crucial impact on the success of learning and many strategies are concerned with judicious discretization (see for instance [FI93, dM93, BG93]) Decision trees are reputed for being both understandable and efficient. As a matter of fact, they produce concise knowledge bases and this knowledge is expressed within the language of the examples ; moreover, the optimality criterion allows for noise filtering. The only heuristics somehow ....

M. Botta and A. Giordana. Smart+ : A multi-strategy learning tool. In Proceedings of IJCAI-93, pages 937--943. Morgan Kaufmann, 1993.

....number handling capabilities of CLP, without requirement for additional background theory. 1 Introduction This paper is devoted to learning from positive and negative examples expressed in first order logic. Many learners have been developed in the field of Inductive Logic Programming (ILP) see [18, 22, 3, 24, 2, 13, 19], among others. However, Logic Programming (LP) and consequently ILP, does not allow for efficient handling of numbers: as emphasized by Saraswat [26] all concepts and operations of interest in [their] underlying domain of computation must be explicitly encoded in the form of Herbrand terms and ....

....work differs from all ILP approaches we are aware of, in that it aims at characterizing all maximally discriminant clauses covering a given example. In contrast, many learners are concerned with finding specific generalizations [18, 24, 20] or one discriminant generalization of an example [22, 3, 19], which can be transformed into an element of G via a post pruning phase. The issue of redundant learning was motivated in [29] redundant learning ensures robust learning when dealing with noisy data, or data that imperfectly encode the underlying problem. The point of noise can be addressed via ....

M. Botta and A. Giordana. Smart+: A multi-strategy learning tool. In Proceedings of IJCAI-93, pages 937--943. Morgan Kaufmann, 1993.

....only, little attention has been payed to numeric features, up to now. The first attempt to deal with continuous attributes in FOL is due to INDUCE with the Closing Interval rule [11] The same solution has been reproposed in other systems [7] inspired to INDUCE. Another proposal has been made in [5], where an algorithm has been proposed in order to learn continuous intervals in SMART . Nevertheless, most algorithms for learning in FOL are not able to directly deal with continuous attributes, but need to transform the data into a discrete representation, with a consequent loss of information ....

M. Botta and A. Giordana. SMART+: A multi-strategy learning tool. In IJCAI-93, Proceedings of the Thirteenth International Joint Conference on Artificial Intelligence, pages 937--943, Chamb'ery, France, 1993.

....organization. Determining explanations for detected execution failures can become very complex when errors are propagated. The proposed approach to modeling errors in terms of taxonomic and causal links aims at handling this complexity. Further information can be found in [2, 32, 8] with [6] describing the basic learning machinery. Transportation and Navigation In addition to actual manufacturing tasks occuring in a factory, transportation of both raw material and fabricated workpieces is an important issue. Driver less transport systems are designed to handle these tasks. However, ....

M. Botta and A. Giordana. SMART+: A multistrategy learning tool. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI '93), Chamberry, France, 1993.

....the incorporation of machine learning paradigms in terms of obtaining a supervision system that is able to learn the necessary knowledge to perform its function. In terms of research this project was very profitable. A number of publications in journals, conferences and workshops can be mentioned [23, 13, 6, 10, 7, 8, 11, 5, 14, 9, 20, 24, 25, 26, 22, 23, 23, 112, 119, 120, 116, 118, 117, 115, 110, 114, 111, 113, 109, 89, 88, 87, 82, 84, 85, 86, 83]. 4 RESULTS 23 Three PhD thesis and one MsC thesis, addressing different areas in the workpackage, are approaching its conclusion. An overview of the main results will now be given. 4.3.1 Results in Assembly Migration from and or integration of legacy systems is one of the most challenging ....

....The best method was then adopted in SKIL. From the supervision point of view, however, the results were not satisfactory, due the low classification accuracies on unseen examples (according to the Leave one out accuracy test) Experiments with other learning systems (namely CART [21] and Smart [20]) were also conducted. Accuracy did not go beyond 90 , and the most typical accuracy results were between 50 and 70 . In robotics applications, information about the status of the system must often be obtained from complex sensors that provide numerical data difficult to analyse. The most obvious ....

M. Botta and A. Giordana. SMART+: A multi-strategy learning tool. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI '93), Chamberry, France, 1993.

....by the tolerance problem. We investigated how much performance degraded when using learnt information on other fixtures. We also showed that it is possible to improve performance when using a new fixture location. The scientific output of this workpackage also can be shown by numerous publications [16, 104, 103, 108, 130, 34, 40, 145, 54, 43, 59, 19, 41, 50, 53, 52, 66, 101, 68, 58, 55, 129, 131, 132, 51, 42, 15, 17], among which several joint publications. 4.3 Results in Assembly and Machining The goal of workpackage WP3 was to enhance supervision architectures in assembly robots and in NC machines by introducing learning capabilities. Activities in the workpackage concerning assembly supervision were ....

M. Botta and A. Giordana. SMART+: a multi-strategy learning tool. In Proc. of the International Joint Conference on Artificial Intelligence, IJCAI-93, Chamb'ery, France, 1993. REFERENCES 35

....The results are reported in table 1. 5.2 Learning a Fuzzy Controller using SMART The synthesis of a Fuzzy Controller from examples has been posed as a problem of learning concept descriptions from examples, which can be solved using a symbolic inductive algorithm. In the specific case, SMART [11] has been used. Coming back to fig. 2, it can be found that the problem is to learn the fuzzy sets of the first layer, the fuzzy sets of the third layer and the set of rules representing the mapping from the first layer to the third one. In the experiments described here, the number and the size ....

M. Botta and A. Giordana. SMART+: A multi-strategy learning tool. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI '93), Chamberry, France, 1993.

....concept to be learned by induction. Then, the problem of learning a continuous control function will be mapped to the problem of learning j H j classification theories: one for each of the linguistic variables H in output to the rule matcher. In the experiment, described here, the system SMART [6] has been used to accomplish the inductive task. In particular, SMART is an inductive system capable of inducing first order logic concept descriptions from examples; the propositional calculus is seen just as a particular case. The inductive process follows a general to specific strategy similar ....

....by an expert and then refined automatically by the system. This feature is useful when the amount of learning events is not large enough to induce a reliable description, or, alternatively, when one wants to force solutions of a particular type. A detailed description of SMART can be found in [6]. A second feature of SMART , which is very useful in the particular task we are facing, is the possibility of dealing with numerical constants by means of parametric predicates . As an example, suppose F x is the force component along the x axis, as described in Figure 10. In SMART , it is ....

M. Botta and A. Giordana. SMART+: a multi-strategy learning tool. In Proc. of the International Joint Conference on Artificial Intelligence, IJCAI-93, Chambery, France, 1993.

....where immediately available to the authors. The first algorithm is the well known CART [3] which can work both for classification and for regression. In both cases the trees produced by CART can be simply translated into rules and then used to construct a network. The second algorithm is SMART [2] which has the peculiarity of combining inductive and deductive techniques. In this way it is possible to force the system to exploit a domain theory even when it is imperfect. Afterwards, the rules learned by CART or SMART are translated into a F RBFN which is finely tuned by performing the ....

M. Botta and A. Giordana. SMART+: A multi-strategy learning tool. In IJCAI93, Proceedings of the Thirteenth International Joint Conference on Artificial Intelligence, volume 2, Chamb'ery, France, 1993.

....hypotheses and their extensions were stored (Bergadano et al. 1988) This feature proved to be essential in learning an industrial troubleshooter (Giordana et al. 1993b) used in field. ML SMART evolved later into two new systems, both including inductive and deductive components: SMART (Botta and Giordana, 1993), which emphasizes handling noise and continuous attributes by exploiting powerful heuristics for controlling the search in the hypothesis space, and WHY (Baroglio et al. 1994; Saitta et al. 1993) which stresses the importance of exploiting background knowledge, in the form of a causal model of ....

.... declarative bias, aimed at reducing the expressive power of the hypothesis language, is the solution most frequently adopted in ILP (Ade et al. 1995) Systems like SMART , instead, have a much weaker declarative bias, but they widely exploits procedural biases in the form of search heuristics (Botta and Giordana, 1993). Increasing the search power has been tried for the first time (in FOL) in the system REGAL, which learns relations via a genetic algorithm (Giordana and Sale, 1992; Giordana and Saitta, 1994; Giordana and Neri, 1996) Genetic search for learning concepts from examples has been used, previously, ....

[Article contains additional citation context not shown here]

Botta, M. and Giordana, A. (1993). SMART+: A multi-strategy learning tool. In IJCAI-93, Proceedings of the Thirteenth International Joint Conference on Artificial Intelligence, pages 937--943, Chamb'ery, France.

No context found.

Botta M., Giordana A.: SMART+: A Multi-Strategy Learning Tool, Proc of the IJCAI-93, 1993

No context found.

Botta, M., Giordana, A. (1993). SMART+ a multi-strategy learning tool. In: Proceedings of the IJCAI'93, Chambery, France, vol. 2.

No context found.

Botta, M., A. Giordana (1993). SMART+: A Multi-Strategy Learning Tool, Proc. IJCAI-93, Chambery, France, pp. 937943.

No context found.

M. Botta, A Giordana, (1993). SMART+ : A MultiStrategy Learning Tool, in IJCAI-93, pp 937-943.

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