真のモデルを含まないパラメトリックモデル族に対するベイズ予測の漸近評価
Bibliographic Information
- Other Title
-
- シン ノ モデル オ フクマナイ パラメトリック モデルゾク ニ タイスル ベイズ ヨソク ノ ゼンキン ヒョウカ
- Asymptotics of Bayesian prediction for misspecified models
- 情報理論
- ジョウホウ リロン
Search this article
Abstract
We consider the sequential prediction problem which is the prediction of the next symbol based on the sequential observation of source symbols. The log loss function in this problem is classified into two types, the instantaneous loss and the cumulative loss. The former is the loss function for the prediction of the only next one symbol. The latter is the sum of the instantaneous loss. We consider the Bayesian prediction for this problem. In Bayesian prediction, it is assumed that the true model lies within a parametrized family of distributions. However, it can be considered that it lies without a parametrized family practically(misspecified models), the true model being unknown. We analyze asymptotics of the cumulative loss for Bayesian prediction under this situation.
Journal
-
- 電子情報通信学会技術研究報告 = IEICE technical report : 信学技報
-
電子情報通信学会技術研究報告 = IEICE technical report : 信学技報 111 (142), 71-76, 2011-07
The Institute of Electronics, Information and Communication Engineers