An agent-based framework for distributed learning

Abstract

In this paper an agent-based distributed learning framework based on data reduction is proposed. Data reduction aims at finding patterns or regularities within certain features, allowing to induce the so-called prototypes which should be retained for further use during the learning process. The considered approach assumes that data reduction through instance and feature selection is carried out independently at each site by a team of agents. To assure obtaining homogenous prototypes the feature selection requires coordination. The proposed approach provides such coordination by collaboration of agents. In the process of data reduction heterogeneous prototypes can be subsequently merged to create a compact representation of the distributed data repositories and, next, based on such a compact representation a selected meta-learning technique can be applied for generating the global classifier. The paper proposes and explains strategies for agent collaboration producing a common set of features and strategies for constructing combiner classifier. Suggested strategies are evaluated experimentally and compared. The paper includes a detailed description of the proposed approaches and a discussion of the computational experiment results.