| 研究課題/領域番号 |
14F04796
|
|---|
| 研究機関 | 東京大学 |
|---|
研究代表者 |
MARTENS Kai 東京大学, カブリ数物連携宇宙研究機構, 准教授 (20535025)
|
|---|
| 研究分担者 |
CALLAND RICHARD 東京大学, カブリ数物連携宇宙研究機構, 外国人特別研究員
|
|---|
| 研究期間 (年度) |
2014-04-25 – 2017-03-31
|
| キーワード | RJMCMC / GPU / CUDA / infinite mixture model / non-parametric model / MCMC |
|---|
| 研究実績の概要 |
The main focus of this period was to work on improving the reconstruction of Super-Kamiokande atmospheric neutrino data. To facilitate this, several computers were purchased using the grant-in-aid research funds, which contain graphics processing units. The improvements to the reconstruction software has 2 main focusses; improving the ring finding algorithm and improving the speed. The ring finding improvement is done via modelling the event with a non-parametric Bayesian mixture model, from which the posterior probability of the model given the data can be computed numerically using a Reversible Jump Markov Chain Monte Carlo. The second task is to improve the speed. Currently the reconstruction takes ～5 minutes to complete per event, and given that the Monte Carlo contains millions of events, this is a huge computational burden. The algorithm is highly parallelizable, and as such I converted most of the charge and time prediction code, along with the likelihood computation to run on GPUs using the CUDA software API. This resulted in a significant speed improvement with negligible sacrifice to numerical precision. The speed improvement is on the order of 50 times faster than the original version.
|
| 現在までの達成度 (区分) |
現在までの達成度 (区分)
3: やや遅れている
理由
The project is a little behind schedule because of unforeseen issues related to the original reconstruction software. It was found that the calculation of time likelihood in the original code contained many approximations, some of which inhibited the new mixture model method. A significant amount of time was used to debug and correct these issues, however the result of this is that the time likelihood has been improved for the original algorithm. Besides this, the new reconstruction method is working and now exhaustive validation studies are ongoing. The GPU acceleration has greatly improved the turnaround time for debugging and developing the code, so progress from this point is expected to be much faster.
|
| 今後の研究の推進方策 |
The plan for FY2016 is to run the new reconstruction algorithm on the atmospheric Monte Carlo and data, to produce new samples which can be used in the oscillation analysis. Using these samples, more detailed sensitivity studies can be undertaken using the analysis framework developed in FY2014. Work is currently ongoing with other groups to improve the systematic uncertainties, so I intend to combine efforts with these groups to come to an acceptable parameterization of uncertainties to be used in the final analysis.
|
