In many cases, performance depends on configurable parameters just as much as the algorithm and data structures used. Therefore tuning of these configurable parameters is needed. Auto-tuning systems are systems that can automatically generate variations of the tuned software, and using a sample run scenario, a derivative free based optimization algorithm, and an evaluation metric, often run-time, it empirically searches, and returns the best configuration found for the tuned system.
The optimization algorithm used by the auto-tuning system, often depends on its own configurable parameters, therefore its performance, both in terms of convergence speed, and quality of the converged value, depend on them as well. I propose an approach to adapt the parameters of the optimization algorithm, used by an auto-tuning system, to the function it tries to optimize (the tuned system performance metrics). This approach assumes no knowledge of the function's behavior and tries to learn and adapt the search algorithm to it, before starting the actual search.
Chair: Dr. Jeff Hollingsworth
Dept rep: Dr. Howard Elman
Member: Dr. Alan Sussman