1.采样方法介绍
近似推断
- 确定性 → V I \to VI →VI
- 随机 → M C M C \to MCMC →MCMC
Monte Carlo Method:基于采样的随机近似方法
推断主要是求后验 P ( z ∣ x ) → P(z \mid x)\to P(z∣x)→后验的期望 E z ∣ x [ f ( z ) ] = ∫ p ( z ∣ x ) f ( z ) d z E_{z \mid x}[f(z)]=\int p(z \mid x)f(z)dz Ez∣x[f(z)]=∫p(z∣x)f(z)dz,积分不好求一般采用数值积分的方式,从 q ( z ∣ x ) q(z \mid x) q(z∣x)中采样:
z ( 1 ) , z ( 2 ) , ⋯ , z ( N ) ∼ p ( z ∣ x ) E z ∣ x [ f ( z ) ] = ∫ p ( z ∣ x ) f ( z ) d z ≈ 1 N ∑ i = 1 N f ( z i ) z^{(1)},z^{(2)},\cdots,z^{(N)} \sim p(z \mid x)\\ E_{z \mid x}[f(z)] =\int p(z \mid x)f(z)dz \approx \frac{1}{N} \sum_{i=1}^{N}f(z_i) z(1),z(2),⋯,z(N)∼p(z∣x)Ez∣x[f(z)]=∫p(z∣x)f(z)dz≈N1i=1∑Nf(zi)
如何从一个复杂的分布采样?
-
概率分布采样
概率密度函数pdf → \to → cdf
u ( i ) ∼ U ( 0 , 1 ) x ( i ) = c d f − 1 ( u ( i ) ) x ( 1 ) , x ( 2 ) , ⋯ , x ( N ) N 个 样 本 u^{(i)} \sim U(0,1)\\ x^{(i)}=cdf^{-1}(u^{(i)})\\ x^{(1)},x^{(2)},\cdots,x^{(N)}N个样本 u(i)∼U(0,1)x(i)=cdf−1(u(i))x(1),x(2),⋯,x(N)N个样本
从U中随机采样作为cdf的逆函数的输入,得到采样点 x ( i ) x^{(i)} x(i)
只有几个简单的pdf能求出对应的cdf
-
拒绝采样
q ( z ) q(z) q(z)为提议分布
∀ z i , M q ( z ( i ) ) ≥ p ( z ( i ) ) \forall z_i,M q(z^{(i)}) \ge p(z^{(i)}) ∀zi,Mq(z(i))≥p(z(i))
定义 α \alpha α为接受率, α = p ( z ( i ) ) M q ( z ( i ) ) , 0 ≤ α ≤ 1 \alpha=\frac{p(z^{(i)})}{M q(z^{(i)})},0 \le \alpha \le 1 α=Mq(z(i))p(z(i)),0≤α≤1,在 q ( z ) q(z) q(z)内接受,否则拒绝。
z ( i ) ∼ q ( z ) u ∼ U ( 0 , 1 ) i f u ≤ α z^{(i)} \sim q(z)\\ u \sim U(0,1)\\ if \ u \le \alpha z(i)∼q(z)u∼U(0,1)if u≤α,接收 z ( i ) z^{(i)} z(i)(绿线)
e l s e else \quad else拒绝(红线)
- Importance Sampling
重要性采样并不是对概率分布采样,它是对概率分布的期望进行采样,引入提议分布 q ( z ) q(z) q(z),从 q ( z ) q(z) q(z)中采样。
E p ( z ) [ f ( z ) ] = ∫ p ( z ) f ( z ) d z = ∫ p ( z ) q ( z ) q ( z ) f ( z ) d z = ∫ f ( z ) p ( z ) q ( z ) q ( z ) d z ≈ 1 N ∑ i = 1 N f ( z ( i ) ) p ( z ( i ) ) q ( z ( i ) ) z ( i ) ∼ q ( z ) , i = 1 , 2 , ⋯ , N \begin{aligned} E_{p(z)}[f(z)] &= \int p(z)f(z)dz \\ &=\int \frac{p(z)}{q(z)}q(z)f(z)dz\\ &=\int f(z)\frac{p(z)}{q(z)}q(z)dz\\ &\approx \frac{1}{N} \sum_{i=1}^{N} f(z^{(i)}) \frac{p(z^{(i)})}{q(z^{(i)})} \\ &z^{(i)} \sim q(z),i=1,2,\cdots,N \end{aligned} Ep(z)[f(z)]=∫p(z)f(z)dz=∫q(z)p(z)q(z)f(z)dz=∫f(z)q(z)p(z)q(z)dz≈N1i=1∑Nf(z(i))q(z(i))p(z(i))z(i)∼q(z),i=1,