Μηχανική μάθηση για το συγκερασμό πολλαπλών, κατανεμημένων ευφυών συστημάτων

Abstract

Nowadays Machine Learning techniques have matured and are embedded in commercial data analysis software packages. Each of the techniques has its advantages and disadvantages and no single technique has excelled the rest for a range of di®erent application domains. Therefore, one issue that arises is the e®ective management of all these methods for the e±cient solution of a speci?c problem. This issue appears in many areas of Arti?cial Intelligence such as Planning, Constraints and Scheduling. This thesis concerns a set of Machine Learning methods for the combination of multiple, distributed intelligent systems with the aim of optimizing the performance on a speci?c problem, along with their implementation using the technology of Web Services. The methods were developed for two categories of intelligent systems: a) classi?ers and b) planners. Speci?cally, a method for the combination of heterogeneous classi?ers was proposed, that results in an increase of accuracy compared to a single classi?er system. The accuracy of the proposed method is analogous and better than the state-of-the-art methods for combining heterogeneous classi?ers, while it has a signi?cantly lower computational cost. It is based on the use of statistical procedures for the evaluation and selection of a subset of algorithms that exhibit the best accuracy. In addition, a method for combining classi?er trained from the data of distributed databases was proposed. The method has the capability of detecting groups of distributed classi?ers with di®erent semantics. It then exploits this information in order to combine classi?ers that belong to the same group. This way the prediction accuracy is increased compared to either a single classi?er, or a global classi?er produced by classical combination methods. At the same time the method o®ers useful knowledge concerning the semantic similarity of distributed data bases. Furthermore, a set of methods was proposed for the selection of either the most suitable parameter con?guration for a planner or the most suitable system from a group of systems for a speci?c problem. Apart from the characteristics of each problem, the methods take into con- sideration the performance requirements of the users or the application domain. Experimental results have shown that the proposed methods lead to much better results compared to the use of a single planning technique or a single planning system. Finally, some of the above methods were implemented into systems using the technology of Web Services, in order to allow the combination of distributed systems. This is especially important in our times, because of the increasing deployment of intelligent systems as Web Services instead of stand-alone software programs. The accomplishment of the combination of distributed intelligent systems requires the availability of publicly acceptable standards that will allow the exchange of data between the distributed software and the execution of their functions. The leading standard today is Web Services.