WEKO3
アイテム
欠測データ補完アルゴリズムと機械学習による太陽系外惑星の全容解明
https://doi.org/10.15102/0002000452
https://doi.org/10.15102/0002000452ced2eb9a-346f-4597-83af-622f600ad02c
| 名前 / ファイル | ライセンス | アクション |
|---|---|---|
LalandeFlorianFulltext.pdf (9.5 MB)
|
|
|
|
LalandeFlorianExamAbstract.pdf (48 KB)
|
|
| Item type | 学位論文 / Thesis or Dissertation(1) | |||||||
|---|---|---|---|---|---|---|---|---|
| 公開日 | 2024-04-15 | |||||||
| タイトル | ||||||||
| タイトル | 欠測データ補完アルゴリズムと機械学習による太陽系外惑星の全容解明 | |||||||
| 言語 | ja | |||||||
| タイトル | ||||||||
| タイトル | Planetary Systems Insights through Numerical Data Imputation Algorithms and Machine Learning | |||||||
| 言語 | en | |||||||
| 言語 | ||||||||
| 言語 | eng | |||||||
| 資源タイプ | ||||||||
| 資源タイプ識別子 | http://purl.org/coar/resource_type/c_db06 | |||||||
| 資源タイプ | doctoral thesis | |||||||
| ID登録 | ||||||||
| ID登録 | 10.15102/0002000452 | |||||||
| ID登録タイプ | JaLC | |||||||
| アクセス権 | ||||||||
| アクセス権 | open access | |||||||
| アクセス権URI | http://purl.org/coar/access_right/c_abf2 | |||||||
| 著者 (英) |
Lalande, Florian
× Lalande, Florian
|
|||||||
| 抄録 | ||||||||
| 内容記述タイプ | Abstract | |||||||
| 内容記述 | Since the first discoveries in the early 1990’s, the number of known exoplanets has exploded to reach over 5,500 as of December 2023. But the recorded information for each planets is sparse, with a lot of missing values, preventing from confidently drawing overarching conclusions. As most traditional data imputation methods provide a point estimate, they fail at capturing the complexity of multimodal data distributions and provide unreliable estimates in scenarios where data exhibits multiple modes. This calls for a new paradigm to model rich or complex numerical datasets. This PhD thesis introduces the kNN×KDE, a numerical imputation tool which combines the flexibility of the k-nearest neighbors (kNN) and the simplicity of Kernel Density Estimation (KDE) to model the multi-dimensional distribution of missing data in datasets characterized by multimodality. This new method is tested against traditional and novel data imputation algorithms, and I show that the kNN×KDE not only provides better estimates, but also facilitates their interpretation. To demonstrate the practical significance of the kNN×KDE, I apply it to the NASA Exoplanet Archive – a dataset riddled with missing values, including both planetary radius and mass, and marked by pronounced multimodality. The analysis of the estimated distributions provided relevant insights into the demographics of the Exoplanet Population, potentially helping future missions to select interesting targets. In addition, this PhD work includes two artificial neural network applications for planetary system analysis: a Convolutional Neural Network (CNN) to predict planetary system stability and a Graph Neural Network (GNN) to rediscover Newton’s Law of Gravitation and attempt to reproduce the scientific discovery of Neptune. Finally, this thesis features a Transformer model for Symbolic Regression applied to 120 real-world physics equations. These additional tools contribute to further characterize planetary systems evolution and understand the limits of Machine Learning for scientific discovery. |
|||||||
| 言語 | en | |||||||
| 口頭試問日 | ||||||||
| 言語 | en | |||||||
| 値 | 2024-02-27 | |||||||
| 学位授与年月日 | ||||||||
| 学位授与年月日 | 2024-03-31 | |||||||
| 学位名 | ||||||||
| 学位名 | Doctor of Philosophy | |||||||
| 学位授与番号 | ||||||||
| 学位授与番号 | 甲第147号 | |||||||
| 学位授与機関 | ||||||||
| 学位授与機関識別子Scheme | kakenhi | |||||||
| 学位授与機関識別子 | 38005 | |||||||
| 学位授与機関名 | Okinawa Institute of Science and Technology Graduate University | |||||||
| 著者版フラグ | ||||||||
| 出版タイプ | VoR | |||||||
| 出版タイプResource | http://purl.org/coar/version/c_970fb48d4fbd8a85 | |||||||
| 権利 | ||||||||
| 権利情報 | © 2024 The Author. | |||||||
