borg logo

an algorithm portfolio toolkit and research project at the University of Texas at Austin

Visualizations generated

Exploring solver performance data

The borg project has also developed an experimental web-based tool for visualizing data on solver performance, especially the data collected by large-scale evaluations like the SAT competitions. The section to the left lists links to visualizations of some of these competitions. A modern browser is required, and Google Chrome is strongly recommended.

This tool visualizes the computational cost, and the success or failure, of runs made by different solvers on collections of instances of some decision problem. It provides two views into these data: the raw data can be explored in the "Tabular View", and a similarity-based projection of all instances can be analyzed in the "Projection View". A probabilistic model—the same model that underlies the borg portfolio solver—is used to quantify instances' similarity, defined as the KL divergence between their solver runtime distributions as predicted by the model.

The source is on github. We welcome contributions and feedback.