Generating attributed variability models for transfer learning

Modern software systems often provide configuration options for customizing of the system's functional and non-functional properties, such as response time and energy consumption. The valid configurations of a software system are commonly documented in a variability model. Supporting the optimization of a system's non-functional properties, variability models have been extended with attributes that represent the influence of one or multiple options on a property. The concrete values of attributes are typically determined only in a single environment (e.g., for a specific software version, a certain workload, and a specific hardware setup) and are applicable only for this context. Changing the environment, attribute values need to be updated. Instead of determining all attributes from scratch with new measurements, recent approaches rely on transfer learning to reduce the effort of obtaining new attribute values. However, the development and evaluation of new transfer-learning techniques requires extensive measurements by themselves, which often is prohibitively costly. To support research in this area, we propose an approach to synthesize realistic attributed variability models from a base model. This way, we can support research and validation of novel transfer-learning techniques for configurable software systems. We use a genetic algorithm to vary attribute values. Combined with a declarative objective function, we search a changed attributed variability model that keeps some key characteristics while mimicking realistic changes of individual attribute values. We demonstrate the applicability of our approach by replicating the evaluation of an existing transfer-learning technique.