Today’s adaptive software systems (i) are often highly configurable product lines, exhibiting hundreds of potentially conflicting configuration options; (ii) are context dependent, forcing the system to reconfigure to ever-changing contextual situations at runtime; (iii) need to fulfill context-dependent performance goals by optimizing measurable nonfunctional properties. Usually, a large number of consistent configurations exists for a given context, and each consistent configuration may perform differently with regard to the current context and performance goal(s). Therefore, it is crucial to consider nonfunctional properties for identifying an appropriate configuration. Existing black-box approaches for estimating the performance of configurations provide no means for determining context-sensitive reconfiguration decisions at runtime that are both consistent and optimal, and hardly allow for combining multiple context-dependent quality goals. In this paper, we propose a comprehensive approach based on Dynamic Software Product Lines (DSPL) for obtaining consistent and optimal reconfiguration decisions. We use training data obtained from simulations to learn performance-influence models. A novel integrated runtime representation captures both consistency properties and the learned performance-influence models. Our solution provides the flexibility to define multiple context-dependent performance goals. We have implemented our approach as a standalone component. Based on an Internet-of-Things case study using adaptive wireless sensor networks, we evaluate our approach with regard to effectiveness, efficiency, and applicability.