Artificial Intelligence systems must be adaptive if they are to operate effectively in complex, real-world domains. In many real-world situations, knowledge must be frequently and sporadically accessed; it may be impossible to make large numbers of observations prior to decision making. This proposal addresses the problem of incremental learning, which assumes that environmental observations are assimilated as they become available. A desirable characteristic of such systems is that they maintain an accurate knowledge base that can be efficiently accessed and updated. However, efficient assimilation of new observations may require that we be satisfied with a knowledge base of lower overall `quality'. Thus, incremental learning systems trade quality against cost, hopefully not to the significant detriment of the former. We begin by reviewing a number of incremental concept learning systems with special attention to their management of the cost/quality tradeoff. From these, we abstract control strategies and knowledge representation schemes that support efficient, but accurate incremental induction. Our analysis motivates proposals for improved incremental learning systems and draws from the literature of cognitive psychology, as well as belief revision. Our primary goal is to investigate domain-independent learning methods, but we have concrete plans to test our systems in areas such as intelligent tutoring, game-playing, and expert knowledge acquisition.