Online Decision Making Based on Random Sampling
| Project/Area Number |
15H02667
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (B)
|
| Allocation Type | Single-year Grants |
|---|
| Section | 一般 |
|---|
| Research Field |
Theory of informatics
|
|---|
| Research Institution | Kyushu University |
|---|
Principal Investigator |
Takimoto Eiji 九州大学, システム情報科学研究院, 教授 (50236395)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
畑埜 晃平 九州大学, 基幹教育院, 准教授 (60404026)
|
|---|
| Project Period (FY) |
2015-04-01 – 2019-03-31
|
| Project Status |
Completed (Fiscal Year 2018)
|
| Budget Amount *help |
¥12,350,000 (Direct Cost: ¥9,500,000、Indirect Cost: ¥2,850,000)
Fiscal Year 2018: ¥2,600,000 (Direct Cost: ¥2,000,000、Indirect Cost: ¥600,000)
Fiscal Year 2017: ¥2,860,000 (Direct Cost: ¥2,200,000、Indirect Cost: ¥660,000)
Fiscal Year 2016: ¥2,600,000 (Direct Cost: ¥2,000,000、Indirect Cost: ¥600,000)
Fiscal Year 2015: ¥4,290,000 (Direct Cost: ¥3,300,000、Indirect Cost: ¥990,000)
|
|---|
| Keywords | 計算学習理論 / オンライン予測 / オンラインアルゴリズム / 組合せ最適化 / ブースティング / 時系列データ / ZDD / 競合比解析 / メトリカルタスクシステム問題 |
|---|
| Outline of Final Research Achievements |
We proposed high accuracy and efficient algorithms for various problems of online decision making, machine learning and combinatorial optimization, by using the methodology of random sampling. Among them we give two main achievements below. (1) We showed that the metrical task system problem over a combinatorial decision space can be reduced to the problem of random sampling over the decision space, and using the reduction we succeeded to give, for the first time, high accuracy and efficient algorithms for various combinatorial decision spaces such as paths of a graph. (2) We proposed a new approach toward large scale machine learning: compress the given training data into a ZDD, and simulate Boosting efficiently over the ZDD (without decompression), based on an online prediction method over the set of accepting paths as the decision space.
|
| Academic Significance and Societal Importance of the Research Achievements |
メトリカルタスクシステム（MTS）問題は，組合せ集合を決定空間とする場合，効率の良いアルゴリズムの統一的で有用な設計法は知られていなかった．本研究は，この問題に対し，ランダムサンプリングの設計問題に落とし込むという，世界初の統一的・現実的な設計指針を与えたといえる．また，大規模機械学習の問題に対し，従来は，確率的勾配降下法などのランダムサンプリングに基づくアプローチが主流であったが，本研究では，圧縮データ上の機械学習という新しいアプローチを提案している．計算時間とメモリ効率の向上を同時に達成する画期的な手法であるだけでなく，圧縮データ上の最適化という新たな研究の方向性を示唆している．
|
Report
(5 results)
Research Products
(52 results)
-
-
-
-
-
-
-
-
-
-
[Journal Article] Boosting over non-deterministic ZDDs2018
Author(s)
Takahiro Fujita, Kohei Hatano and Eiji Takimoto
-
Journal Title
Proceedings of the 12th International Conference of International Workshop on Algorithms and Computation(WALCOM 2018)
Volume: LNCS 10755
Pages: 195-206
DOI
NAID
ISBN
9783319751719, 9783319751726
Related Report
Peer Reviewed / Open Access
-
-
-
-
-
-
[Journal Article] Minimax Fixed-Design Linear Regression2015
Author(s)
Peter Bartlett, Wouter Koolen, Alan Malek, Eiji Takimoto, Manfred Warmuth
-
Journal Title
Proceedings of The 28th Conference on Learning Theory (COLT 2015), JMLR: Workshop and Conference Proceedings
Volume: 40
Pages: 226-239
Related Report
Peer Reviewed / Open Access / Int'l Joint Research
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
[Presentation] Boosting over ZDDs2017
Author(s)
Takahiro Fujita, Kohei Hatano, Eiji Takimoto
Organizer
The 20th Korea-Japan Joint Workshop on Algorithms and Computation
Related Report
Int'l Joint Research
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
