This paper explores the use of learning as a practical tool in problem solving. The idea that learning should and eventually will be a vital component of most Artificial Intelligence programs is pursued. Current techniques in learning systems are compared. A detailed discussion of the problems of representing, modifying, and creating heuristics is given. Some of the questions asked (and answered) in the paper are: (1) how does the choice of representation affect the potential for learning?, (2) what techniques have been used to date and how do they compare?, i.e. first-order predicate calculus vs. production rules vs. Winston's representation, and (3) exactly how are heuristics modified in the existing systems and what do these techniques have in common? A discussion of the credit assignment problem as it relates to learning under the various schemes of representation is also presented.