Matthieu Bloch
Tuesday, November 8, 2022
\[ \begin{aligned} \hat{x}_{i|i} &= \hat{x}_{i|i-1}+K_{f,i}(y_i-h(\hat{x}_{i|i-1}))\\ K_{f,i} &= P_{i|i-1}H_x(\hat{x}_{i-1|i-1})^\intercal(H_x(\hat{x}_{i-1|i-1})P_{i|i-1}H_x(\hat{x}_{i-1|i-1})^\intercal+R_i)^{-1}\\ P_{i|i}&= P_{i|i-1}-K_{f,i}H_x(\hat{x}_{i-1|i-1}) P_{i|i-1} \end{aligned} \]
EKF is based on Taylor series expansions of \(f\) and \(h\) around current state estimate
Advantages of EKF
Limitations of EKF
Linearization assumes that all second and higher order terms in Taylor series expansion are negligible
Possible to extend EKF approach to higher order terms
Matching mean and covariance of a Gaussian \[ \begin{aligned} \mathbf{c}_1[\calS,p_x(\bfx)]&=\sum_{i=0}^p w^{(i)}\bfx^{(i)} - \bar{\bfx}\\ \mathbf{c}_2[\calS,p_x(\bfx)]&=\sum_{i=0}^p w^{(i)}(\bfx^{(i)}-\bar{\bfx})(\bfx^{(i)}-\bar{\bfx})^\intercal - \Sigma_x\\ \end{aligned} \]
Solution: set of points that lie on the \(\sqrt{N_x}\) covariance contour
Propagate sigma points through non-linear transform
In general, \(p(y_{1:T}|\vecx^{(i)})\) easier to evaluate (measurement model)
Use Bayes' rule \[ p(\vecx^{(i)}|\vecy_{1:T}) = \frac{p(y_{1:T}|\vecx^{(i)})p(\vecx^{(i)})}{\int p(y_{1:T}|\vecx)p(\vecx)d\vecx} \] and note the denominator is still hard to compute!
Importance sampling to the rescue \[ \E{g(\vecx)|\vecy_{1:T}} \approx \sum_{i=1}^n \tilde{w}^{(i)} g(\vecx^{(i)})\textsf{ where } \tilde{w}^{(i)} \eqdef \frac{\frac{p(\vecy_{1:T}|\vecx^{(i)})p(\vecx^{(i)})}{\pi(\vecx^{(i)}|\vecy_{1:T})}}{\sum_{j=1}^n\frac{p(\vecy_{1:T}|\vecx_j)p(\vecx_j)}{\pi(\vecx_j|\vecy_{1:T})}} \]
Sequential importance sampling is a variation that can be used for probabilistic state-space models with \[ \vecx_k\sim p(\vecx_k|\vecx_{k-1})\qquad \vecy_k\sim p(\vecy_k|\vecx_k) \]
Idea