p********0 发帖数: 186 | 1 For the state-space linear dynamic system,
y_t+1 = A_t * x_t + R
x_t+1 = B_t * x_t + Q
I read some paper about using EM - maximum likelihood to estimate the
parameter A and B
In simple case assume A_t = A_t-1 = ... A_1, we can maximize \hat{x_t} = E[x
_t | Y_1,,,n]
How do we handle the situaion A_t change over time, like ARMA(1, 1) case,
just use fewer observation Y_t-1, Y_t to estimat the A_t/B_t ??? | p********0 发帖数: 186 | 2 Ding,
I may not be clear about my question, let me rephrase it.
For time series data, state space model, if the parameter is not static,
change over time
X_t = F_t * X_t-1 + R.
Y_t = H_t * X_t
What's the best way for finding the online parameter?
I found resurve Least Sqaure Error/Gaussian Newton method/Kalman filter all
can be used to do the online parameter estimate, what's the best method?
Anyone has experience?
[x
【在 p********0 的大作中提到】 : For the state-space linear dynamic system, : y_t+1 = A_t * x_t + R : x_t+1 = B_t * x_t + Q : I read some paper about using EM - maximum likelihood to estimate the : parameter A and B : In simple case assume A_t = A_t-1 = ... A_1, we can maximize \hat{x_t} = E[x : _t | Y_1,,,n] : How do we handle the situaion A_t change over time, like ARMA(1, 1) case, : just use fewer observation Y_t-1, Y_t to estimat the A_t/B_t ???
| g****t 发帖数: 31659 | 3 你得知道F_t大概变化的快慢。对F_t一无所知,那就没办法辨识。
例如200个点之内F可看作常数,那你就每200个点作一次least square。
all
【在 p********0 的大作中提到】 : Ding, : I may not be clear about my question, let me rephrase it. : For time series data, state space model, if the parameter is not static, : change over time : X_t = F_t * X_t-1 + R. : Y_t = H_t * X_t : What's the best way for finding the online parameter? : I found resurve Least Sqaure Error/Gaussian Newton method/Kalman filter all : can be used to do the online parameter estimate, what's the best method? : Anyone has experience?
|
|