-
- 山西 健司
- NEC C&Cメディア研究所
書誌事項
- タイトル別名
-
- Extended Stochastic Complexity and Informatical Learning Theory
- カクチョウガタ カクリツテキ コンプレキシティ ト ジョウホウロンテキ ガクシ
この論文をさがす
抄録
Rissanen has introduced stochastic complexity to define the amount of information in a given data sequence relative to a given hypothesis class of probability densities, where the information is measured in terms of a logarithmic loss associated with universal data compression. This paper proposes the notion of extended stochastic complerxity (ESC) by generalizing Rissanen's stochastic complexity into the decision-theoretic setting where a general real-valued function is used as a hypothesis and a general loss function is used as a distortion measure. We thereby demonstrate the effectiveness of ESC in design and analysis of learning algorithms in sequential prediction and function-estimation scenarios. As an application of ESC to sequential prediction, this paper shows that a sequential realization of ESC produces a sequential prediction algorithm called the aggregating strategy, for which the worst-case cumulative prediction loss is asymptotically minimal. As an application of ESC to function-estimation, this paper shows that a batch-approximation of ESC induces a batchlearning algorithm called the minimum L-complexity algorithm (MLC), for which an upper bound on the statistical risk is least to date. Through ESC we give a unifying view of designing the most effective learning algorithms in fundamental learning issues.
収録刊行物
-
- 応用数理
-
応用数理 8 (3), 188-203, 1998
一般社団法人 日本応用数理学会
- Tweet
詳細情報 詳細情報について
-
- CRID
- 1390282680743536512
-
- NII論文ID
- 110007390781
- 10012150161
-
- NII書誌ID
- AN10288886
-
- ISSN
- 09172270
- 24321982
-
- NDL書誌ID
- 4559324
-
- 本文言語コード
- ja
-
- データソース種別
-
- JaLC
- NDL
- CiNii Articles
-
- 抄録ライセンスフラグ
- 使用不可