In this paper we extend previous work on approximate planning in large stochastic domains by adding the ability to plan in automaticailygenerated abstract world views. The dynamics of the domain are represented compositionally using a Bayesian network. Sensitivity analysis is performed on the network to identify the aspects of the world upon which success is most highly dependent. An abstract world model is constructed by including only the most relevant aspects of the world. The world view can be refined over time, making the overall planner behave in most cases like an anytime algorithm. This paper is a preliminary report on this ongoing work.