This paper presents a model-based black-box testing approach for dynamically adaptive systems. Behavioral models of such systems are formally specified using timed automata. With the aim of obtaining the new test suite and avoiding its regeneration in a cost effective manner, we propose a selective test generation approach. The latter comprises essentially three modules: (1) a model differencing module that detects similarities and differences between the initial and the evolved behavioral models, (2) an old test classification module that identifies reusable and retestable tests from the old test suite, and finally (3) a test generation module that generates new tests covering new behaviors and adapts old tests that failed during animation. To show its efficiency, the proposed technique is illustrated through the Toast application and compared to the classical Regenerate All and Retest All approaches.