Presentation | 2018-11-05 [Poster Presentation] Proposal of Hyperparameter Optimization Framework Using a Non-Stationary Multi-Armed Bandit Algorithm Kenshi Abe, Masahiro Nomura, |
---|---|
PDF Download Page | PDF download Page Link |
Abstract(in Japanese) | (See Japanese page) |
Abstract(in English) | Hyperparameter optimization problem is an important problem that appears in areas such as machine learning. Hyperparameter is classified into continuous variable or categorical variable. It is conceivable that an optimal value of the continuous variable is different per the categorical variable. So continuous variable should be represented by the form that ties to categorical variable. Also in the optimization for continuous variable, the optimal algorithm is clearly different per problems. Therefore, we need a framework that have the form that continuous variable ties to categorical variable and the mechanism that an optimization algorithm can be replaced with other one for continuous variable per problems. However, in case of using the framework, it is difficult to decide to optimize for which categorical variable and continuous variable that ties to it because the performance of categorical variable changes depending on the search situation for continuous variable. In this paper, we consider that the above problem arise from the non-stationarity of the distribution of the evaluation value, and we propose the framework HOTS using Thompson Sampling to resolve the problem. On the benchmark problem that the distribution of a evaluation value is non-stationary and on the hyperparameter optimization problem for Deep Neural Network, our experimental results show that HOTS achieves better performance than the comparison methods. |
Keyword(in Japanese) | (See Japanese page) |
Keyword(in English) | Hyperparameter Optimization / Non-stationary Bandit / Thompson Sampling / Deep Neural Networks |
Paper # | IBISML2018-62 |
Date of Issue | 2018-10-29 (IBISML) |
Conference Information | |
Committee | IBISML |
---|---|
Conference Date | 2018/11/5(3days) |
Place (in Japanese) | (See Japanese page) |
Place (in English) | Hokkaido Citizens Activites Center (Kaderu 2.7) |
Topics (in Japanese) | (See Japanese page) |
Topics (in English) | Information-Based Induction Science Workshop (IBIS2018) |
Chair | Hisashi Kashima(Kyoto Univ.) |
Vice Chair | Masashi Sugiyama(Univ. of Tokyo) / Koji Tsuda(Univ. of Tokyo) |
Secretary | Masashi Sugiyama(Nagoya Inst. of Tech.) / Koji Tsuda(AIST) |
Assistant | Tomoharu Iwata(NTT) / Shigeyuki Oba(Kyoto Univ.) |
Paper Information | |
Registration To | Technical Committee on Infomation-Based Induction Sciences and Machine Learning |
---|---|
Language | JPN |
Title (in Japanese) | (See Japanese page) |
Sub Title (in Japanese) | (See Japanese page) |
Title (in English) | [Poster Presentation] Proposal of Hyperparameter Optimization Framework Using a Non-Stationary Multi-Armed Bandit Algorithm |
Sub Title (in English) | |
Keyword(1) | Hyperparameter Optimization |
Keyword(2) | Non-stationary Bandit |
Keyword(3) | Thompson Sampling |
Keyword(4) | Deep Neural Networks |
1st Author's Name | Kenshi Abe |
1st Author's Affiliation | CyberAgent, Inc.(CA) |
2nd Author's Name | Masahiro Nomura |
2nd Author's Affiliation | CyberAgent, Inc.(CA) |
Date | 2018-11-05 |
Paper # | IBISML2018-62 |
Volume (vol) | vol.118 |
Number (no) | IBISML-284 |
Page | pp.pp.135-142(IBISML), |
#Pages | 8 |
Date of Issue | 2018-10-29 (IBISML) |