Increasingly, machine learning is entertained as a mechanism for improving the efficiency of planning systems. Research in this area has generated an impressive battery of techniques and a growing body of empirical successes. Unfortunately the formal properties of these systems are not well understood. This is highlighted by a growing corpus of demonstrations where learning actually degrades planning performance. In this paper we view learning to plan as a search problem. Learning is seen as a transformational process where a planner is tailored to a particular domain and problem distribution. To accomplish this task, learning systems draw from a vocabulary of transformation operators such as macro-operators or control rules. These "learning operators" define a space of possible transformations through which a system must search for an efficient planner. This study shows that the complexity of this search precludes a general solution and can only be approached via simplifications. (Frequently unarticulated commitments which underlie current learning approaches are illustrated.) These simplifications improve learning efficiency but not without tradeoffs. In some cases these tradeoffs result in less than optimal behavior. In others, they produce planners which become worse through learning. It is hoped that by articulating these commitments we can better understand their ramifications can be better understood. Finally, a particular learning technique--COMPOSER--is discussed which explicitly utilizes these simplifications to ensure performance improvements with reasonable efficiency. (Contains 34 references.) (Author/ALF)