Model-Driven Optimization: Towards Performance-Enhancing Low-Level Encodings

Paper i proceeding, 2023

In Model-Driven Optimisation, meta-heuristic opti-mization algorithms are applied to models to solve optimization problems. A meta-model is used to describe a modelling language which defines the search space. Exploration operators (e.g., mutation) are usually expressed as model transformations. During the search space exploration, transformations as well as model copying can become a performance bottleneck, significantly slowing down performance. In this paper, as a first step towards solving this issue, we contribute a low-level encoding of models that does not replace, but compliments them. The encoding stores information about the mutable parts of the model in a way that is inexpensive to change and copy, whereas other operations (e.g., querying of non-mutable parts) are still performed on the actual model. We include a formal framework for expressing what such an encoding looks like, together with an implementation on top of MDEOptimiser, an existing tool for Model-Driven Optimization. In a performance evaluation on two scenarios, we find improved performance in one, and new, clearly identified performance challenges in a second scenario.