def likelihood(self, y, x, observation_point, sigma = 2):
"""
likelihood関数
p(y|x) ~ exp(|g_inverse(y)-x|**2 / simga**2)で定義
yは観測値,xは隠れマルコフ変数,sigmaはデータからつくる
"""
x_from_y = self.g_inverse(y, observation_point)
v = mymath.vector()
distance = range(len(y))
for i in range(len(y)):
distance[i] = x[i] - x_from_y[i]
norm = math.pow(v.norm(distance), 2)
return math.exp(-norm / math.pow(sigma, 2))
def pf_sir_step(self, X, W, y, sigma, date):
"""One Step of Sampling Importance Resampling for Particle Filter
Parameters
----------
X : array of [float|array]
状態 List of state set
W : array of float
重み List of weight
y : float or array
観測 Observation set
Returns
-------
X_resampled : array of [float|array]
次の状態 List updated state
W_resampled : array of float
次の重み List updated weight
"""
# パーティクルの数 num of particles
N = len(X)
# 次元数
D = len(X[0])
# 初期化
X_predicted = [[1. for j in range(D)] for i in range(N)]
W_updated = [1. for i in range(N)]
# 観測点の取得
obs = self.select_observation_point(date)
# 推定 prediction
for i in range(N):
X_predicted[i] = self.state_transition(X[i])
# 更新 update
for i in range(N):
W_updated[i] = self.importance_sampling(W[i], X_predicted[i], y, obs, sigma)
# 正規化 normalization
v = mymath.vector()
W_updated = v.normalization(W_updated)
# リサンプリング re-sampling (if necessary)
X_resampled, W_resampled = self.resampling(X_predicted, W_updated)
return X_resampled, W_resampled
