Abstract: In order to enable ``anywhere, anytime'' computing, pervasive
systems must autonomously adapt at runtime. The use of dynamic software
product lines has emerged as a promising paradigm where well established
variability management techniques are leveraged at runtime to describe
evolution strategies and adaptation scenarios in terms of combinations of
features. In order to identify the optimal target configuration of the
system under certain circumstances, most existing approaches generate the
set of valid combinations of features and return the best one. Obviously,
while such approaches are well suited to small systems with a reduced number
of configurations, they fail in the case of large modern pervasive systems
because the generation/evaluation of all valid combinations is very costly
in terms of resources and time consumption. In the present article, we
introduce a new scalable, evolutionary-based approach to runtime adaptation
of pervasive systems. To this end, we define the concept of transitive
dependency between features and we exploit it to fasten the generation of
the optimal configuration of the system. We evaluate the scalability of our
proposal by reporting experimental results that show that our genetic
algorithm converges in up to 90% less time than the one from the
literature while preserving the exploration capabilities and solutions
quality. Finally, we illustrate our proposal on the smart homes use case.
