| No | 180271 |
|---|---|
| 標題（和） | On the convergence of iterative estimation algorithms operating on graphical models |
| 標題（英） | On the convergence of iterative estimation algorithms operating on graphical models |
| 研究会名（和） | 信号処理, 回路とシステム, 通信方式 |
| 研究会名（英） | Signal Processing, Circuits and Systems, Communication Systems |
| 開催年月日 | 2007-03-05 |
| 終了年月日 | 2007-03-06 |
| 会議種別コード | 5 |
| 共催団体名（和） | |
| 資料番号 | CAS2006-81, SIP2006-182, CS2006-98 |
| 抄録（和） | In recent years, graphical models have become a useful tool to derive estimation and detection algorithms. Perhaps the most prominent example in digital communications is the probabilistic decoding of error correcting codes (e.g., LDPC codes, turbo-codes, etc.) by means of the sum-product algorithm (``belief propagation\'\') applied on a graphical model (i.e., Tanner graph or factor graph) of the error-correcting code. Graphical models have also been used in the context of channel estimation, e.g., for phase, frequency, and timing estimation. \r\nIn this paper, we investigate the convergence properties of various estimation algorithms that operate on graphical models: (generalized) expectation maximization, gradient methods, and cyclic maximization. First we will review those algorithms in the standard formulation, then we consider various extensions in the context of graphical models. We will elaborate on the fixed points, asymptotic rate of convergence (in terms of information matrices), and local stability of those algorithms. |
| 抄録（英） | In recent years, graphical models have become a useful tool to derive estimation and detection algorithms. Perhaps the most prominent example in digital communications is the probabilistic decoding of error correcting codes (e.g., LDPC codes, turbo-codes, etc.) by means of the sum-product algorithm (``belief propagation\'\') applied on a graphical model (i.e., Tanner graph or factor graph) of the error-correcting code. Graphical models have also been used in the context of channel estimation, e.g., for phase, frequency, and timing estimation. \r\nIn this paper, we investigate the convergence properties of various estimation algorithms that operate on graphical models: (generalized) expectation maximization, gradient methods, and cyclic maximization. First we will review those algorithms in the standard formulation, then we consider various extensions in the context of graphical models. We will elaborate on the fixed points, asymptotic rate of convergence (in terms of information matrices), and local stability of those algorithms. |
| 収録資料名（和） | 電子情報通信学会技術研究報告 |
| 収録資料の巻号 | Vol.106, No.567,569,571 |
| ページ開始 | 23 |
| ページ終了 | 28 |
| キーワード（和） | fixed points,convergence rate,gradient methods,Hessian matrix,graphical model,cyclic maximization,stability,expectation maximization |
| キーワード（英） | fixed points,convergence rate,gradient methods,Hessian matrix,graphical model,cyclic maximization,stability,expectation maximization |
| 本文の言語 | ENG |
| 著者（和） | Justin Dauwels |
| 著者（ヨミ） | |
| 著者（英） | Justin Dauwels |
| 所属機関（和） | Amari Research Unit, RIKEN Brain Science Institute |
| 所属機関（英） | Amari Research Unit, RIKEN Brain Science Institute |