In machine learning there is considerable interest in techniques which improve planning ability. Initial investigations have identified a wide variety of techniques to address this issue. Progress has been hampered by the utility problem, a basic tradeoff between the benefit of learned knowledge and the cost to locate and apply relevant knowledge. In this paper we describe the COMPOSER system which embodies a probabilistic solution to the utility problem. We outline the statistical foundations of our approach and compare it against four other approaches which appear in the literature.