Towards accuracy-adaptive simulation for efficient design-space optimization

Designing complex embedded systems is a challenge, which benefits from using models for various tasks. Among them, an important one is finding a good (or the best) system design alternative for a given set of changeable parameters. Such decisions are usually based on non-functional properties which can be quantitatively expressed and evaluated using a model. Automatic optimization would be advantageous, but one of its problems is the enormous computational effort to evaluate the model for a large number of parameter sets. In this paper a new idea to speed up indirect optimization is proposed, which makes use of an integration of optimization control and underlying performance evaluation technique. Moreover, we propose to adaptively use refinements or abstractions of complex hierarchical models to control the tradeoff between model complexity and result accuracy of a performance evaluation, and demonstrate this dependency with an example.