Alternative model representations and computing capacity: implications for model management

Recent research on model management systems (MMS) recognizes the importance of considering potential algorithmic performance in the selection of an appropriate model to solve a real-world problem. Model selection, as typically viewed in the literature however, is the process of selecting from among alternative model classes, rather than from alternative mathematical representations of the same model class. In this paper, we take up this subtler aspect of model selection, and provide tangible evidence that shows how just changing the representation of a model can have a dramatic impact on algorithmic performance. Using problem decomposition and distributed processing, we conduct a series of computational experiments to study the interrelationships between model representation, computing capacity, and algorithmic performance. We discuss potential implications of our results for improving MMS design and address a key prerequisite for the enhanced design, by proposing and validating an approach for solution time prediction.