20
Mixture Kalman Filters by Rong Chen & Jun Liu Presented by Yusong Miao Dec. 10, 2003

Mixture Kalman Filters by Rong Chen & Jun Liu Presented by Yusong Miao Dec. 10, 2003

Embed Size (px)

Citation preview

Page 1: Mixture Kalman Filters by Rong Chen & Jun Liu Presented by Yusong Miao Dec. 10, 2003

Mixture Kalman Filtersby Rong Chen & Jun Liu

Presented by Yusong MiaoDec. 10, 2003

Page 2: Mixture Kalman Filters by Rong Chen & Jun Liu Presented by Yusong Miao Dec. 10, 2003

Structure

Introduce the originals of the problem Mixture Kalman Filters (MKF) model

setup and method Two related extended models w/

examples Applications to show the advantages of

MKF Conclusions

Page 3: Mixture Kalman Filters by Rong Chen & Jun Liu Presented by Yusong Miao Dec. 10, 2003

Original of the problem Interest in on-line estimation and prediction of the

dynamic changing system. (Hidden pattern along observations)

Kalman filter (1960) technique can OK Gaussian linear system.

How about non-linear & non-Gaussian system? ------Sequential Monte Carlo approach including: Bootstrip filter / practical filter; Sequential imputation; From on now, call it as “Monte Carlo filters” ------Mixed kalman filter, the role of this paper. Will see the comparisons.

Page 4: Mixture Kalman Filters by Rong Chen & Jun Liu Presented by Yusong Miao Dec. 10, 2003

Original of the problem

Before we start MKF, recall the task:

Page 5: Mixture Kalman Filters by Rong Chen & Jun Liu Presented by Yusong Miao Dec. 10, 2003

MKF model setup Conditional dynamic linear model (CDLM):

Given trajectory of an indicator variable, the system is Gaussian & linear-- can derive a MC filter focusing on attention on the space of indicator variables, named Mixed Kalman Filter.

Page 6: Mixture Kalman Filters by Rong Chen & Jun Liu Presented by Yusong Miao Dec. 10, 2003

MKF model setup Example 1: A special CDLM (Linear system with non-

Gaussian errors) as the follow:

In the CDLM system MKF is more sophisticated, outperform other methods (i.e.. bootstrap filter).

Use a mixture of Gaussian distribution to estimate the target posterior distribution.

Page 7: Mixture Kalman Filters by Rong Chen & Jun Liu Presented by Yusong Miao Dec. 10, 2003

MKF model setup The method of MKF

Use a weighted sample of the indicators:

To represent the distribution p(Λt|yt) a random mixture of Gaussian distribution:

Can approximate the target distribution p(xt|yt).

Page 8: Mixture Kalman Filters by Rong Chen & Jun Liu Presented by Yusong Miao Dec. 10, 2003

MKF model setup Algorithm (updating weights)---If you are interested in:

Page 9: Mixture Kalman Filters by Rong Chen & Jun Liu Presented by Yusong Miao Dec. 10, 2003

Extended MKF with examples

Beside the situation of CDLM, there is a extended one called partial conditional dynamic linear models (PCDLM).

PCDLM are interested in non-linear component of state variables.

No absolute distinction between CDLM and PCDLM.

Page 10: Mixture Kalman Filters by Rong Chen & Jun Liu Presented by Yusong Miao Dec. 10, 2003

Extended MKF with examples

Example: Fading channel modeling system (mobile communication channel can be modeled as Rayleigh flat fading

channels)

Page 11: Mixture Kalman Filters by Rong Chen & Jun Liu Presented by Yusong Miao Dec. 10, 2003

Extended MKF with examples

Example: Blind deconvolution digital communication system

Where St is a discrete process taking vales on a known set S. It is to be estimated from the observed signals {y1,…,yt}, without knowing the channel coefficients θi.

There two examples can be solved by extended MKF called EMKF.

Page 12: Mixture Kalman Filters by Rong Chen & Jun Liu Presented by Yusong Miao Dec. 10, 2003

Extended MKF with examples Why EMKF? It can deal with as many linear and

Gaussian components from systems as possible. P(xt1,xt2|yt)=P(xt1|xt2,y)*P(xt2|yt), Monte Carlo

approximation of P(xt2|yt) and an Gaussian conditional distribution p(xt1|xt2,y).

Need to generate discrete samples in the joint space of the indicators and the non-linear state components.

Page 13: Mixture Kalman Filters by Rong Chen & Jun Liu Presented by Yusong Miao Dec. 10, 2003

Extended MKF with examples Algorithm (updating

weights)

Page 14: Mixture Kalman Filters by Rong Chen & Jun Liu Presented by Yusong Miao Dec. 10, 2003

Application of MKF-Target tracking

Situation setup:

The tracking errors (differences between the estimated and true target location) are generated and compared with other methods.

Page 15: Mixture Kalman Filters by Rong Chen & Jun Liu Presented by Yusong Miao Dec. 10, 2003

Application of MKF-Target tracking

The result proves the advantage of MKF:

Page 16: Mixture Kalman Filters by Rong Chen & Jun Liu Presented by Yusong Miao Dec. 10, 2003

Application of MKF-2 D target’s position

A 2-D target’s position is sampled every T=10s. We know the movement and velocities on both x and y directions. Use MKF to simulate the results and compare them with the actual data.

Model setup:

10

01,

10

01

50

05

,

10

01

50

05

,0010

0001,

1000

0100

10010

01001

22vw VWFGH

Page 17: Mixture Kalman Filters by Rong Chen & Jun Liu Presented by Yusong Miao Dec. 10, 2003

Application of MKF-2 D target’s position

The result proves the advantage of MKF:

Better than the tradition way done by Bar-Shalom & Fortmann

Page 18: Mixture Kalman Filters by Rong Chen & Jun Liu Presented by Yusong Miao Dec. 10, 2003

Other applications of MKF

There are still several other applications with brief introduction can be found on the paper.

Random (non-Gaussian) accelerated target (no clutter).

Digital signal extraction in fading channels. They are both improved under MKF approach

comparing with traditional Monte Carlo approach.

Page 19: Mixture Kalman Filters by Rong Chen & Jun Liu Presented by Yusong Miao Dec. 10, 2003

Conclusions MKF can perform real time estimation and prediction

in CDLM situation, which outperform Sequential Monte Carlo approaches.

Similar to EMKF in PCDLM situation. MKF can combine with other Monte Carlo techniques

(Markov chain Monte Carlo updates, delayed estimation, fixed lag filter, etc.) to improve effectiveness.

Sequential Monte Carlo method can be a platform for designing efficient non-linear filtering algorithms.

Page 20: Mixture Kalman Filters by Rong Chen & Jun Liu Presented by Yusong Miao Dec. 10, 2003

Questions & Answers