図1 状態空間モデル |
$\Large \displaystyle x_t=f(x_{t-1},\xi_s,\upsilon_t)\ \ \ \ \ \ \ \ \ $(2.1)「システム方程式」
$\Large \displaystyle y_t=h(x_t,\xi_m,\epsilon_t)\ \ \ \ \ \ \ \ \ \ \ \ $(2,2)「観測方程式」
$\Large \displaystyle x_t \sim \mathbb{P}(x_t|x_{t-1})\ \ \ \ \ \ \ \ \ \ \ \ \ $(2-3)
$\Large \displaystyle y_t \sim \mathbb{P}(y_t|x_{t})\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ $(2-4)
予測分布:$\Large \displaystyle x_{t|t-1} = \mathbb{P}(x_t|y_{1:(t-1)})\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ $(2-5-1)
フィルタ分布:$\Large \displaystyle x_{t|t} = \mathbb{P}(x_t|y_{1:t})\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ $(2-5-2)
平滑化分布:$\Large \displaystyle x_{t|T} = \mathbb{P}(x_t|y_{1:T})\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ $(2-5-3)
$\Large \displaystyle p(x_t|y_{1:t})\simeq \sum_{i=1}^{M}\frac{w_t^i}{\sum_{i=1}^{M}w_t^i}\delta(x_t-x_{t|t-1}^i)$
$\Large \displaystyle x_t=F_t x_{t-1}+G_t v_t\ \ \ \ \ $(1)「システムモデル」
$\Large \displaystyle y_t=H_t x_{t}+w_t\ \ \ \ \ \ \ \ \ \ \ \ $(2)「観測モデル」
$\Large \displaystyle \{ x_{t_0}^{(n)}\}_{n\in [1,...,N]}$
$\Large \displaystyle \forall n\ x_i^{(n)}\sim p(x_t^{(n)}|x_{t-1}^{(n)})$
$\Large \displaystyle \mu_t^{(n)}=p(y_t|x_t^{(n)})$
$\Large \displaystyle \tilde{\mu}_t^{(n)}=\frac{\mu_t^{(n)}}{\sum_{n=1}^N\mu_t^{(n)}}$
$\Large \displaystyle \forall n\ c_t^{(n)}\sim Categorical(c_t^{(n)}|)$
$\Large \displaystyle x_t^{(n)}=c_t^{(n)^T} \begin{pmatrix} x_t^{(1)} \\ \vdots \\ x_t^{(N)} \end{pmatrix} $
$\large \displaystyle \large 1:Algorithm Particle_filter(X_{t-1},u_t, z_t):\\ \large 2:\ \ \ \bar{X_t}=X_t = 0\\ \large 3:\ \ \ for\ M=1\ to\ M\ do\\ \large 4:\ \ \ \ \ \ sample\ x_t^{[m]}\sim p(x_t|u_t, x_{t-1}^{[m]})\\ \large 5:\ \ \ \ \ \ w_t^{[m]}=p(z_t|x_{t}^{[m]})\\ \large 6:\ \ \ \ \ \ \bar{X_t}=\bar{X_t}+\lt x_t^{[m]},w_t^{[m]}\gt\\ \large 7:\ \ \ endfor\\ \large 8:\ \ \ for\ M=1\ to\ M\ do\\ \large 9:\ \ \ \ \ \ draw\ i\ with \ probalility\propto w_t^{[i]}\\ \large 10:\ \ \ \ add\ x_t^{[i]}\ to\ X_t\\ \large 11:\ endfor\\ \large 12:\ return\ X_t $
非線形な変換を適用したガウス分布 |
$\Large \displaystyle \mu= \begin{bmatrix} 0 \\ 0 \end{bmatrix}, \Sigma= \begin{bmatrix} 32 & 15 \\ 15 & 40 \end{bmatrix} $
$\Large \displaystyle x=x+y$
$\Large \displaystyle y=0.1x^2+y^2$
粒子フィルタのアルゴリズムを一般的に説明すると次のようになる: