Some Insights into the Behavior of Long-term Learning in Soar

Much of the work in machine learning has focused on situations in which there are distinct training and testing phases. However, the recent increase in interest in real-world systems involving scaling up has led to an increased attention to “anytime learning”, i.e., systems in which learning mechanisms are always active. This places different demands on machine learning techniques and raises a different set of questions than the more traditional train/test work. As a starting point we have instrumented Soar to allow us to collect data and characterize the long-term behavior of learning in Soar. In this paper we describe our approach and provide a set of initial results.