題 目：Statistical Surrogates: How Statistics May Impact
High Performance Computing
主講人：王偉仲教授（台灣大學數學系）
時 間：101年4月13日(星期五)下午14:00-14:50
地 點：交大綜合一館427室
Abstract
Fast evolving computing technologies have kept bringing excellent
opportunities to process a great amount of data. These technologies have
also kept inducing novel and exciting scientific discoveries. However,
it is quite a challenge to design efficient algorithms on these powerful
computers due to their architecture complexities and the lack of
equation-based performance models. To optimize the performance of computer
algorithms or computer codes, it is common we have to tune the algorithms
or codes via limited data. As such tasks can be modeled as data-driven
optimization problems, performance surrogates constructed by statistical
tools play a critical role in these computer experiments. We illustrate such
approach by showing how design and analysis of computer experiments (DACE)
and expected improvement (EI) can be applied to highly time consuming
numerical simulation for three-dimensional photonic crystals shape
optimization. The potential of statistical surrogates is far beyond the
aforementioned description. Taking a new medical imaging methodology and a
software auto-tuning framework as examples, we scratch some thoughts on a
promising perspective that statistical surrogates is a key player for the
latest and the next generation computers.
--
※ 發信站: 批踢踢實業坊(ptt.cc)
◆ From: 140.113.114.163