Planning and learning in domains providing little feedback

Letters, and speech acts in general, provide little feedback about their success or failure, which makes it difficult for generators to improve their planning knowledge. Informal observation indicates that humans may learn better writing techniques through discovery rather than failure, however. I present a bottom-up, case-based planner that was developed to overcome some shortcomings of classical planners in planning speech acts, and I discuss how this planning technique addresses the problem of learning better techniques as well.