自動チューニングのための数理基盤技術(<特集>数値計算のための自動チューニング)

概要

論文の詳細を見る
This paper introduces the concepts of automatic tuning, and discusses mathematical aspects and methods of automatic tuning. First, the concepts of automatic tuning are introduced as automation of manual tuning, in order to clarify the goal automatic tuning research is seeking. Next, mathematical aspects of automatic tuning are investigated and analyzed. Based on that framework of automatic tuning analysis, one case of our previous works is re-analyzed and discussed. Our mathematical method, a Bayes-based sequential experimental design with linear models is revisited as well. Next, this paper newly proposes "tuning meter," which quantitatively shows the progress of automatic tuning process. As a Bayesian method, our method provides quantitative estimate of current belief, which can be utilized in a tuning meter. Last, the concept of "infinite dilution" is explained as a method to make existing methods of optimization and machine learning available in online automatic tuning.
2010-09-24