Foundations of Intelligent Systems

An increasing number of services are appearing both within agent communities and as Web Services on the World Wide Web. As these services proliferate, humans and agents need to be able to find, select, understand and invoke these services. Today, services (e.g. travel services, book selling services, stock reporting services etc) are discovered and invoked manually by human users. In the near future, such service discovery and use will be mediated by agents acting on behalf of their human users. Such use of agent technology will be the next Web revolution. Instead of populated with humanreadable documents, the Web will be populated with Agent-Mediated Services. For this to be accomplished, the Web must become agent-understandable, i.e. allow for semantic annotation of content. Up to now, this vision has been conceived and pursued mainly in academia and research labs. However, recent industrial interest in such services, and the availability of tools to enable service automation (e.g. UDDI, WSDL, X-lang, WSFL, e-speak, .NET etc) holds the promise of fast progress in the automation in the Web Services area. Agent Mediated discovery, selection, execution and monitoring of Web Services will be a crucial test of Agent Technology. Agent Mediated Web Services is a confluence of Agent Technology and the Semantic Web. In order to enable stable and scalable Agent Mediated Services, a widely used, widely accessible and extensible Multiagent (MAS) infrastructure is crucial. Part of this infrastructure should be languages for semantic annotation of content as well as for describing services, so that they can be discovered, invoked and composed. DAML-S is such a language for semantic descriptions of services. Another part of the MAS infrastructure should define communication and interoperability of agents. Various standards bodies (e.g. FIPA) are attempting to define standards for various aspects of MAS infrastructure, such as Agent Communications Languages.. However, there is no coherent account of what constitutes a MAS infrastructure, what functionality it supports, what characteristics it should have in order to enable various value-added abilities, such as Agent Based Mediation of Services, and what its possible relation with and requirements it may impose on the design and structure of single agents. In this talk, we will present a model of MAS infrastructure, and our implemented RETSINA system that is an example of the general infrastructure model. In addition, we will show how RETSINA implements Agent Mediated Web Services through a variety of tools and mechanisms. Moreover, we will present DAML-S and illustrate its utility in the area of Agent Mediated Services. Improving Classification by Removing or Relabeling Mislabeled Instances Stéphane Lallich, Fabrice Muhlenbach, and Djamel A. Zighed ERIC Laboratory – University of Lyon 2 5, av. Pierre Mendès-France F-69676 BRON Cedex – FRANCE {lallich, fabrice.muhlenbach, zighed}@univ-lyon2.fr Abstract. It is common that a database contains noisy data. An important source of noise consists in mislabeled training instances. We present a new approach that deals with improving classification accuracies in such a case by using a preliminary filtering procedure. An example is suspect when in its neighborhood defined by a geometrical graph the proportion of examples of the same class is not significantly greater than in the whole database. Such suspect examples in the training data can be removed or relabeled. The filtered training set is then provided as input to learning algorithm. Our experiments on ten benchmarks of UCI Machine Learning Repository using 1-NN as the final algorithm show that removing give better results than relabeling. Removing allows maintaining the generalization error rate when we introduce from 0 to 20% of noise on the class, especially when classes are well separable. It is common that a database contains noisy data. An important source of noise consists in mislabeled training instances. We present a new approach that deals with improving classification accuracies in such a case by using a preliminary filtering procedure. An example is suspect when in its neighborhood defined by a geometrical graph the proportion of examples of the same class is not significantly greater than in the whole database. Such suspect examples in the training data can be removed or relabeled. The filtered training set is then provided as input to learning algorithm. Our experiments on ten benchmarks of UCI Machine Learning Repository using 1-NN as the final algorithm show that removing give better results than relabeling. Removing allows maintaining the generalization error rate when we introduce from 0 to 20% of noise on the class, especially when classes are well separable.