ELA—A new Approach for Learning Agents

摘要

In this paper we discuss a new incremental learning approach used to implement adaptive behavior in autonomous agents. Adaptive agents must increase their performance based on experience using some learning approach. Often, incremental learning techniques like memory-based reasoning (MBR) are used. However, traditional MBR algorithms require an adequate (generally complex) measure of similarity, need much data and spend much time for computing similarities between examples. Such problems are unacceptable for autonomous agents that live in very dynamic environments, because they have little time to make decisions. Our approach does not use similarity measures between examples, classifies examples very fast and can compact data. We represent data as a concept graph (CG), each node representing a partition of the data. We propose an algorithm that uses the partitions to classify new examples. We compare our results with other techniques and conclude that the method performs quite well. Finally, we apply the approach to an application of adaptive agents for personalizing web search.