Angaben zur Quelle [Bearbeiten]
Autor | Tao Chen, Julian Morris, Elaine Martin |
Titel | Probability Density Estimation via Infinite Gaussian Mixture Model: Application to Statistical Process Monitoring |
Zeitschrift | Journal of the Royal Statistical Society. Series C. Applied Statistics |
Ort | Oxford |
Verlag | Wiley-Blackwell |
Jahr | 2006 |
Nummer | 55 |
Seiten | 699-715 |
DOI | 10.1.1.122.8834 |
URL | http://epubs.surrey.ac.uk/6482/2/tchen06-jrscc.pdf (preprint) |
Literaturverz. |
no |
Fußnoten | no |
Fragmente | 4 |
[1.] Rh/Fragment 144 33 - Diskussion Zuletzt bearbeitet: 2012-08-06 16:18:12 Hindemith | Chen et al. 2006, Fragment, Gesichtet, Rh, SMWFragment, Schutzlevel sysop, Verschleierung |
|
|
Untersuchte Arbeit: Seite: 144, Zeilen: 33-35 |
Quelle: Chen et al. 2006 Seite(n): 4, Zeilen: 17-21 |
---|---|
Given the set y the classical approach to estimate the parameters (μl, Ωl), is to maximize the likelihood by using the EM algorithm (Dempster et al. (1977)). The EM algorithm guarantees convergence to a local maximum, with the quality of the maximum being heavily [dependent on the random initialization of the algorithm.] | Given a set of training data with N observations, x={x1,...,xN}, the classical approach to estimating the Gaussian mixture model parameters, (μ,τ,π), is to maximize the likelihood using the expectation-maximization (EM) algorithm (Dempster et al., 1977). The EM algorithm guarantees convergence to a local maximum, with the quality of the maximum being heavily dependant [sic] on the random initialization of the algorithm. |
Eine Quellenangabe fehlt. Die Übernahme setzt sich auf der nächsten Seite fort: Rh/Fragment_145_01 |
|
[2.] Rh/Fragment 145 01 - Diskussion Zuletzt bearbeitet: 2012-08-06 16:18:23 Hindemith | Chen et al. 2006, Fragment, Gesichtet, Rh, SMWFragment, Schutzlevel sysop, Verschleierung |
|
|
Untersuchte Arbeit: Seite: 145, Zeilen: 1-8 |
Quelle: Chen et al. 2006 Seite(n): 4, Zeilen: 22-28 |
---|---|
Alternatively, a Bayesian approach can be used to combine the prior distribution for the parameters and the likelihood, resulting in a joint posterior distribution:
[EQUATION, equivalent to equation in the source] (5.5.2) However, the joint posterior takes a highly complicated form. Thus, it is generally not feasible to perform analytical inference based on the above posterior distribution. The MCMC approaches have typically been used to calculate the joint posterior and, of the approaches that have been proposed in the literature, Gibbs sampling is suitable for mixture models. |
Alternatively, a Bayesian approach can be used to combine the prior distribution for the parameters and the likelihood, resulting in a joint posterior distribution:
[EQUATION] (3) However the joint posterior takes a highly complicated form. Thus it is generally not feasible to perform any analytical inference based on the above posterior distribution. MCMC approaches have typically been used to calculate the joint posterior and of the approaches proposed in the literature, Gibbs sampling is suitable for mixture models |
Keine Quellenangabe vorhanden. Die oben nicht dokumentierte Gleichung ist in der Dissertation equivalent zur Gleichung in der Quelle. Variablen und Verteilungen haben nur andere Namen. |
|
[3.] Rh/Fragment 148 23 - Diskussion Zuletzt bearbeitet: 2014-01-07 15:29:45 Schumann | Chen et al. 2006, Fragment, Gesichtet, Rh, SMWFragment, Schutzlevel sysop, Verschleierung |
|
|
Untersuchte Arbeit: Seite: 148, Zeilen: 23-27, 29-32 |
Quelle: Chen et al. 2006 Seite(n): 7, 8, Zeilen: 7: 1-5; 8: 14-17 |
---|---|
Let k- denote the number of represented mixtures. For represented mixtures, the previously derived conditional posteriors of (μl, Ωl) in (5.4.4) and (5.4.5) still hold. In contrast, in the absence of training data, the parameters in unrepresented mixtures are solely determined by their priors f(μl | ml, Σ / κ ) and f(Ω | 㯑,ψ-1). Thus the inference of the indicators, ci, must incorporate the effect of infinite mixtures. [...]
The conditional posteriors of (μl, Ωl)are Gaussian and Wishart distributions respectively, from which samples can be generated by using standard procedures. The sampling of the indicators requires the evaluation of the integral in equation (5.5.7), which is only analytically feasible if the conjugate prior is used. |
Let krep denote the number of represented mixtures. For represented mixtures, the previously derived conditional posteriors of μj and τj still hold (Eq. (4) and (5)). On the other hand, in the absence of training data, the parameters in unrepresented mixtures are
solely determined by their priors (p(μj | λ , γ ) and p (τj | β , ω )), Thus the inference of the indicators, c, has to incorporate the effect of infinite mixtures. [Seite 8] The conditional posteriors of μj and τj are Gaussian and Gamma distributions respectively, from which samples can be generated using standard procedures. The sampling of the indicators requires the evaluation of the integral in Eq. (12), which is only analytically feasible if the conjugate prior is used [...]. |
Ein Verweis auf die Quelle fehlt, obwohl es sich hier um eine fast wörtliche Übernahme handelt (nur die Notation wurde angepasst). |
|
[4.] Rh/Fragment 150 11 - Diskussion Zuletzt bearbeitet: 2012-07-30 23:32:08 Hindemith | Chen et al. 2006, Fragment, Gesichtet, Rh, SMWFragment, Schutzlevel sysop, Verschleierung |
|
|
Untersuchte Arbeit: Seite: 150, Zeilen: 10-18 |
Quelle: Chen et al. 2006 Seite(n): 8 (preprint), Zeilen: 26-36 |
---|---|
5.5.5. Prediction. The calculation of the predictive probability of new data will be averaged over a number of MCMC samples, which are selected from those samples where the algorithm tends to stabilize. Stabilization will be assessed heuristically based on the value of the log-likelihood. Additionally to eliminate the auto-correlation, one sample will be selected from each consecutive set of 10 iterations. For a particular MCMC sample, the predictive probability is attained from two components: the represented and the unrepresented mixtures. In a similar manner to that adopted in the sampling stage, the probability from unrepresented mixtures will be approximated by a finite mixture of Gaussians, whose parameters (μl, Ωl) are drawn from the prior. | 2.4 Prediction
The calculation of the predictive probability of new data will be averaged over a number of MCMC samples, which are selected from those where the algorithm tends to stabilize. Stabilization will be assessed heuristically based on the value of the log-likelihood. Additionally to eliminate the auto-correlation, one sample will be selected from each consecutive set of 10 iterations. For a particular MCMC sample, the predictive probability is attained from two components: the represented and the unrepresented mixtures. In a similar manner to that adopted in the sampling stage, the probability from unrepresented mixtures will be approximated by a finite mixture of Gaussians, whose parameters, (μj, τj) are drawn from the prior. |
identisch; ohne Kennzeichnung und ohne Quellenangabe. |
|