This research seeks to build a theoretical framework for when to learn. Machine learning, which studies the automatic acquisition of new concepts, has largely focussed on developing mechanisms for learning. Questions such as "How can an AI computer system come to form a general concept in response to a relatively few observed training examples?" and "What prior knowledge is required to support such concept acquisition?" have dominated the field. However, it has become clear that the unbridled acquisition of concepts (even correct concepts) can degrade the performance of the AI system that they are intended to help. In sophisticated applications, such as arise in AI planning, a performance penalty for learning can be the rule rather than the exception. Thus, machine learning mechanisms cannot be successful unless suitably restrained. Unfortunately, the judgement of when to learn is itself complex. The benefit of a new concept depends on features of the performance system as well as the machine learning mechanism. It can also be strongly influenced by the expected profile of tasks to be given to the performance system and by the concepts that have previously been acquired. The proposed research is intended to produce a fist theory of rational learning. The learning component of an AI system is a rational learner if the AI system's behavior is guaranteed, on average, to be improved by the acquisition of any new concepts. Such a theory must provide a general framework for the interactions between a machine learning system, the performance system that employs the acquired concepts, and characteristics of the estimated distribution of tasks. The research will explore both empirical and analytic approaches to estimating expected concept utility. The area of planning will serve as a vehicle for the work, but the results will likely be applicable to other performance systems. Explanation-based methods form the nucleus of learning methods since it, more than the inductive approach, seems liable to the phenomenon of detrimental learning. The benefits of the research will include an enumeration and taxonomy of the types of rational learning. It will shed light on which machine learning mechanisms may be most useful to which type of AI systems, and represents some first steps towards making machine learning a service area for the rest of AI.
