This paper presents an approach to developing self-adaptive systems that takes the end users’ preferences into account for adaptation planning, while tolerating incomplete and conflicting adaptation goals. The approach transforms adaptation goals, together with the run-time model that describes current system contexts and configurations, into a constraint satisfaction problem. From that, it diagnoses the conflicting adaptation goals to ignore, and determines the required re-configuration that satisfies all remaining goals. If users do not agree with the solution, they can revise some configuration values. The approach records their preferences embedded in the revisions by tuning the weights of existing goals, so that subsequent adaptation results will be closer to the users’ preferences. The experiments on a medium-sized simulated smart home system show that the approach is effective and scalable.