Using data mining algorithms for statistical learning of a software agent

In many applications software agents are supposed to show adaptive behaviour and learning capabilities in information rich environments. On the other hand agents are often expected to be resource-bounded systems, which do not utilize much memory, disk space and CPU time. In this paper we present a novel framework for incremental, statistical learning, attempting to satisfy both requirements. The new method, called APS, runs in a cycle including such phases as: storing observations in a history, rule discovery using data mining algorithms, and knowledge base maintenance. Once processed, the old facts are removed from the history and in every subsequent learning run only the recent portion of observations is analysed in search of new rules. This approach can substantially save disk space and processing time as compared to batch learning methods.