Victor Eijkhout,
Texas Advanced Computing Center, The University of Texas at Austin, USA

　 "A Self-adapting System for Linear Solver Selection"

Abstract:
We present research in progress on the SALSA (Self-Adapting
Large-scale Solver Architure) system which is designed to recommend
the optimal solver for a given problem.

A self-adapting system integrates several components: feature
extraction of problem characteristics, a database for storage of
features and runtime results, a modeller that builds recommendation
strategies from the information in the database, and a runtime
recommender. The SALSA system is well under way to releasing libraries
for each of these components. We will describe the general ideas of
the system, and its implementation.

The recommendation strategies being explored in the SALSA system are
based on Bayesian classification, Boosting, and Reinforcement
learning. We report encouraging results from applying these techniques
to the problem of solver recommendation.