Meta learning in hybrid multi-agent systems designed for data mining

Introduction Discovering patterns in data usually requires deeper understanding of both data and data mining methods. We have developed a multi-agent system (MAS) which is able to meta learn over different configurations of computational methods (such as multilayer perceptron or RBF network, embedded in agents), to gather experience, and to utilize results of previous tasks in order to find the best possible method even for new datasets. The combination of search heuristics for parameter spaces of computational methods and ontology based matching of datasets should partially replace human expert experience. The multi-agent-based approach brings in many advantages to the complex task of meta learning. The main contribution lies in its parallel and distributed nature and the easy extensibility of MAS, which in our case is assured by use of the structured ontology language, by following international standards of agents’ communication and also by the general description of our MAS using roles [1]. The structure of our MAS is depicted in Fig.1.