Characteristics of Long-term Learning in Soar and its Application to the Utility Problem

Much of the work in machine learning has focused on demonstrating the efficacy of learning techniques using training and testing phases. On-line learning over the long term places different demands on symbolic machine learning techniques and raises a different set of questions for symbolic learning than for empirical learning. We have instrumented Soar to collect data and characterize the long-term learning behavior of Soar and demonstrate an effective approach to the utility problem. In this paper we describe our approach and provide results.