Lecture2: Bellman Equation
State value
考虑grid-world的单步过程:
S t → A t R t + 1 , S t + 1 S_t \xrightarrow[]{A_t} R_{t + 1}, S_{t + 1} StAtRt+1,St+1
- t t t, t + 1 t + 1 t+1:时间戳
- S t S_t St:时间 t t t时所处的state
- A t A_t At:在state S t S_t St时采取的action
- R t + 1 R_{t + 1} Rt+1:在采取 action A t A_t At 之后获得的reward
- S t + 1 S_{t + 1} St+1:在采取 action A t A_t At 之后,state S t S_t St转移后的state
通过概率分布对以上变量的动作进行描述:
- S t → A t S_t \rightarrow A_t St→At: π ( A t = a ∣ S t = s ) \pi (A_t = a | S_t = s) π(At=a∣St=s)
- S t , A t → R t + 1 S_t, A_t \rightarrow R_{t + 1} St,At→Rt+1: p ( R t + 1 = r ∣ S t = s , A t = a ) p(R_{t + 1} =r | S_t = s, A_t = a) p(Rt+1=r∣St=s,At=a)
- S t , A t → S t + 1 S_t, A_t \rightarrow S_{t + 1} St,At→St+1: p ( S t + 1 = s ′ ∣ S t = s , A t = a ) p(S_{t + 1} = s' | S_t = s, A_t = a) p(St+1=s′∣St=s,At=a)
考虑grid-world的多步(multi-step)trajectory:
S t → A t R t + 1 , S t + 1 → A t + 1 R t + 2 , S t + 2 → A t + 2 R t + 3 . . . S_t \xrightarrow[]{A_t} R_{t + 1}, S_{t + 1} \xrightarrow[]{A_{t + 1}} R_{t + 2}, S_{t + 2} \xrightarrow[]{A_{t + 2}} R_{t + 3}... StAtRt+1,St+1At+1Rt+2,St+2At+2Rt+3...
其discounted return为:
G t = R t + 1 + γ R t + 2 + γ 2 R t + 3 + . . . G_t = R_{t + 1} + \gamma R_{t + 2} + \gamma^2 R_{t + 3} + ... Gt=Rt+1+γRt+2+γ2Rt+3+...
- γ ∈ [ 0 , 1 ) \gamma \in [0, 1) γ∈[0,1)是折扣率(discount rate)
- 当 R t + 1 , R t + 2 , . . . R_{t + 1}, R_{t + 2}, ... Rt+1,Rt+2,...是随机变量时, G t G_t Gt也是随机变量
G t G_t Gt的期望(expectation; expected value; mean)被定义为state-value function或state value。
v π ( s ) = E [ G t ∣ S t = s ] v_{\pi}(s) = \mathbb{E}[G_t | S_t = s] vπ(s)=E[Gt∣St=s]
- v π ( s ) v_{\pi}(s) vπ(s)是state s s s的函数,是state从 s s s 起始的条件期望。
- v π ( s ) v_{\pi}(s) vπ(s)基于policy π \pi π,对于不同的policy,state value可能会不同
- 其代表了一个state的“价值”。 如果state value越大,代表policy就越好,因为可以获得更大的累积奖励(cumulative rewards)。
注意区分state value和return: state value是从某个state开始可以获得的所有可能return的平均值。如果每一个 π ( a ∣ s ) , p ( r ∣ s , a ) , p ( s ′ ∣ s , a ) \pi(a | s), p(r | s, a), p(s' | s, a) π(a∣s),p(r∣s,a),p(s′∣s,a)是确定的,那么state value和return是相等的。
例:
计算三个样例的state value:
v π 1 ( s 1 ) = 0 + γ 1 + γ 2 1 + ⋯ = γ ( 1 + γ + γ 2 + ⋯ ) = γ 1 − γ v_{\pi_1}(s_1) = 0 + \gamma 1 + \gamma^21 + \cdots = \gamma(1 + \gamma + \gamma^2 + \cdots) = \frac{\gamma}{1 - \gamma} vπ1(s1)=0+γ1+γ21+⋯=γ(1+γ+γ2+⋯)=1−γγ
v π 2 ( s 1 ) = − 1 + γ 1 + γ 2 1 + ⋯ = − 1 + γ ( 1 + γ + γ 2 + ⋯ ) = − 1 + γ 1 − γ v_{\pi_2}(s_1) = -1 + \gamma1 + \gamma^21 + \cdots = -1 + \gamma(1 + \gamma + \gamma^2 + \cdots) = -1 + \frac{\gamma}{1 - \gamma} vπ2(s1)=−1+γ1+γ21+⋯=−1+γ(1+γ+γ2+⋯)=−1+1−γγ
v π 3 ( s 1 ) = 0.5 ( − 1 + γ 1 − γ ) + 0.5 ( γ 1 − γ ) = − 0.5 + γ 1 − γ v_{\pi_3}(s_1) = 0.5(-1 + \frac{\gamma}{1 - \gamma}) + 0.5(\frac{\gamma}{1 - \gamma}) = -0.5 + \frac{\gamma}{1 - \gamma} vπ3(s1)=0.5(−1+1−γγ)+0.5(1−γγ)=−0.5+1−γγ
Bellman equation: Derivation
贝尔曼方程描述了所有state值之间的关系。
考虑一个随机的trajectory:
S t → A t R t + 1 , S t + 1 → A t + 1 R t + 2 , S t + 2 → A t + 2 R t + 3 , … S_t \xrightarrow[]{A_t} R_{t + 1}, S_{t + 1} \xrightarrow[]{A_{t+1}} R_{t + 2}, S_{t + 2} \xrightarrow[]{A_{t+2}} R_{t + 3}, \dots StAtRt+1,St+1At+1Rt+2,St+2At+2Rt+3,…
其return G t G_t Gt可以被计算为:
G t = R t + 1 + γ R t + 2 + γ 2 R t + 3 + … = R t + 1 + γ ( R t + 2 + γ R t + 3 + … ) = R t + 1 + γ G t + 1 \begin{align*} G_t &= R_{t + 1} + \gamma R_{t + 2} + \gamma^2 R_{t + 3} + \dots\\ &= R_{t + 1} + \gamma(R_{t + 2} + \gamma R_{t + 3} + \dots)\\ &= R_{t + 1} + \gamma G_{t+1} \end{align*} Gt=Rt+1+γRt+2+γ2Rt+3+…=Rt+1+γ(Rt+2+γRt+3+…)=Rt+1+γGt+1
其state value可以计算为:
v π ( s ) = E [ G t ∣ S t = s ] = E [ R t + 1 + γ G t + 1 ∣ S t = s ] = E [ R t + 1 ∣ S t = s ] + γ E [ G t + 1 ∣ S t = s ] \begin{align*} v_{\pi}(s) &= \mathbb{E}[G_t | S_t = s] \\ & = \mathbb{E}[R_{t + 1} + \gamma G_{t + 1} | S_t = s]\\ &= \mathbb{E}[R_{t + 1} | S_t = s] + \gamma \mathbb{E}[G_{t + 1} | S_t = s] \end{align*} vπ(s)=E[Gt∣St=s]=E[Rt+1+γGt+1∣St=s]=E[Rt+1∣St=s]+γE[Gt+1∣St=s]
对于第一项:
E [ R t + 1 ∣ S t = s ] = ∑ a π ( a ∣ s ) E [ R t + 1 ∣ S t = s , A t = a ] = ∑ a π ( a ∣ s ) ∑ r p ( r ∣ s , a ) r \begin{align*} \mathbb{E}[R_{t + 1} | S_t = s] &= \sum_a \pi(a | s) \mathbb{E}[R_{t + 1} | S_t = s, A_t = a] \\ & = \sum_a \pi(a | s)\sum_rp(r | s, a)r \end{align*} E[Rt+1∣St=s]=a∑π(a∣s)E[Rt+1∣St=s,At=a]=a∑π(a∣s)r∑p(r∣s,a)r
这是瞬时reward的期望。
对于第二项:
E [ G t + 1 ∣ S t = s ] = ∑ s ′ E [ G t + 1 ∣ S t = s , S t + 1 = s ′ ] p ( s ′ ∣ s ) = ∑ s ′ E [ G t + 1 ∣ S t + 1 = s ′ ] p ( s ′ ∣ s ) = ∑ s ′ v π ( s ′ ) p ( s ′ ∣ s ) = ∑ s ′ v π ( s ′ ) ∑ a p ( s ′ ∣ s , a ) π ( a ∣ s ) \begin{align*} \mathbb{E}[G_{t + 1} | S_t = s] &= \sum_{s'} \mathbb{E}[G_{t + 1} | S_t = s, S_{t + 1} = s']p(s' | s)\\ &= \sum_{s'}\mathbb{E}[G_{t + 1} | S_{t + 1} = s']p(s' | s)\\ &= \sum_{s'} v_{\pi}(s')p(s' |s )\\ &= \sum_{s'} v_{\pi}(s') \sum_a p(s' | s, a)\pi(a | s) \end{align*} E[Gt+1∣St=s]=s′∑E[Gt+1∣St=s,St+1=s′]p(s′∣s)=s′∑E[Gt+1∣St+1=s′]p(s′∣s)=s′∑vπ(s′)p(s′∣s)=s′∑vπ(s′)a∑p(s′∣s,a)π(a∣s)
这是未来reward的期望
因此,可以得到:
v π ( s ) = E [ R t + 1 ∣ S t = s ] + γ E [ G t + 1 ∣ S t = s ] = ∑ a π ( a ∣ s ) ∑ r p ( r ∣ s , a ) r + γ ∑ s ′ v π ( s ′ ) ∑ a p ( s ′ ∣ s , a ) π ( a ∣ s ) = ∑ a π ( a ∣ s ) [ ∑ r p ( r ∣ s , a ) r + γ ∑ s ′ p ( s ′ ∣ s , a ) v π ( s ′ ) ] , ∀ s ∈ S \begin{align*} v_{\pi}(s) &= \mathbb{E}[R_{t + 1} | S_t = s] + \gamma \mathbb{E}[G_{t + 1} | S_t = s]\\ &= \sum_a \pi(a | s)\sum_rp(r | s, a)r + \gamma \sum_{s'} v_{\pi}(s') \sum_a p(s' | s, a)\pi(a | s) \\ &= \sum_a \pi(a | s) \left[ \sum_r p(r | s, a)r + \gamma \sum_{s'}p(s' | s, a) v_{\pi}(s') \right], \;\;\; \forall s \in S \end{align*} vπ(s)=E[Rt+1∣St=s]+γE[Gt+1∣St=s]=a∑π(a∣s)r∑p(r∣s,a)r+γs′∑vπ(s′)a∑p(s′∣s,a)π(a∣s)=a∑π(a∣s)[r∑p(r∣s,a)r+γs′∑p(s′∣s,a)vπ(s′)],∀s∈S
- v π ( s ) v_{\pi}(s) vπ(s)和 π ( s ′ ) \pi(s') π(s′)是需要被计算的state value,可以采用bootstrapping。
- π ( a ∣ s ) \pi(a | s) π(a∣s)是给定的policy,可以通过策略评估(policy evaluation)进行求解。
- p ( r ∣ s , a ) p(r | s, a) p(r∣s,a)和 p ( s ′ ∣ s , a ) p(s' | s, a) p(s′∣s,a)代表动态模型,分为known和unknown。
- 上式叫做贝尔曼等式(Bellman equation),其描述了不同state之间state-value function的关系。
- Bellman equation包含两个部分,瞬时奖励(immediate reward)和未来奖励(future reward)。
例:
对于action:
若policy为:
首先写Bellman equation:
v π ( s ) = ∑ a π ( a ∣ s ) [ ∑ r p ( r ∣ s , a ) r + γ ∑ s ′ p ( s ′ ∣ s , a ) v π ( s ′ ) ] v_{\pi}(s) = \sum_a \pi(a | s) \left[ \sum_r p(r | s, a)r + \gamma \sum_{s'}p(s' | s, a) v_{\pi}(s') \right] vπ(s)=a∑π(a∣s)[r∑p(r∣s,a)r+γs′∑p(s′∣s,a)vπ(s′)]
计算上式各项:
- π ( a = a 3 ∣ s 1 ) = 1 \pi(a = a_3 | s_1) = 1 π(a=a3∣s1)=1, π ( a ≠ a 3 ∣ s 1 ) = 0 \pi(a \ne a_3 | s_1) = 0 π(a=a3∣s1)=0
- p ( s ′ = s 3 ∣ s 1 , a 3 ) = 1 p(s' = s_3 | s_1, a_3) = 1 p(s′=s3∣s1,a3)=1, p ( s ′ ≠ s 3 ∣ s 1 , a 3 ) = 0 p(s' \ne s_3 | s_1, a_3) = 0 p(s′=s3∣s1,a3)=0
- p ( r = 0 ∣ s 1 , a 3 = 1 ) p(r = 0 | s_1, a_3 = 1) p(r=0∣s1,a3=1), p ( r ≠ 0 ∣ s 1 , a 3 ) = 0 p(r \ne 0 | s_1, a_3) = 0 p(r=0∣s1,a3)=0
替换进Bellman equation得:
v π ( s 1 ) = 0 + γ v π ( s 3 ) v_{\pi}(s_1) = 0 + \gamma v_{\pi}(s_3) vπ(s1)=0+γvπ(s3)
同样的,可以计算:
v π ( s 1 ) = 0 + γ v π ( s 3 ) v π ( s 2 ) = 1 + γ v π ( s 4 ) v π ( s 3 ) = 1 + γ v π ( s 4 ) v π ( s 4 ) = 1 + γ v π ( s 4 ) v_{\pi}(s_1) = 0 + \gamma v_{\pi}(s_3)\\ v_{\pi}(s_2) = 1 + \gamma v_{\pi}(s_4)\\ v_{\pi}(s_3) = 1 + \gamma v_{\pi}(s_4)\\ v_{\pi}(s_4) = 1 + \gamma v_{\pi}(s_4)\\ vπ(s1)=0+γvπ(s3)vπ(s2)=1+γvπ(s4)vπ(s3)=1+γvπ(s4)vπ(s4)=1+γvπ(s4)
对于上式,可以从后往前计算:
v π ( s 4 ) = 1 1 − γ v π ( s 3 ) = 1 1 − γ v π ( s 2 ) = 1 1 − γ v π ( s 1 ) = γ 1 − γ v_{\pi}(s_4) = \frac{1}{1 - \gamma}\\ v_{\pi}(s_3) = \frac{1}{1 - \gamma}\\ v_{\pi}(s_2) = \frac{1}{1 - \gamma}\\ v_{\pi}(s_1) = \frac{\gamma}{1 - \gamma}\\ vπ(s4)=1−γ1vπ(s3)=1−γ1vπ(s2)=1−γ1vπ(s1)=1−γγ
若policy为:
则:
v π ( s 1 ) = 0.5 [ 0 + γ v π ( s 3 ) ] + 0.5 [ − 1 + γ v π ( s 2 ) ] v π ( s 2 ) = 1 + γ v π ( s 4 ) v π ( s 3 ) = 1 + γ v π ( s 4 ) v π ( s 4 ) = 1 + γ v π ( s 4 ) v_{\pi}(s_1) = 0.5[0 + \gamma v_{\pi}(s_3)] + 0.5[-1 + \gamma v_{\pi}(s_2)] \\ v_{\pi}(s_2) = 1 + \gamma v_{\pi}(s_4)\\ v_{\pi}(s_3) = 1 + \gamma v_{\pi}(s_4)\\ v_{\pi}(s_4) = 1 + \gamma v_{\pi}(s_4)\\ vπ(s1)=0.5[0+γvπ(s3)]+0.5[−1+γvπ(s2)]vπ(s2)=1+γvπ(s4)vπ(s3)=1+γvπ(s4)vπ(s4)=1+γvπ(s4)
从后往前算:
v π ( s 4 ) = 1 1 − γ v π ( s 3 ) = 1 1 − γ v π ( s 2 ) = 1 1 − γ v π ( s 1 ) = 0.5 [ 0 + γ v π ( s 3 ) ] + 0.5 [ − 1 + γ v π ( s 2 ) ] = − 0.5 + γ 1 − γ v_{\pi}(s_4) = \frac{1}{1 - \gamma} \\ v_{\pi}(s_3) = \frac{1}{1 - \gamma} \\ v_{\pi}(s_2) = \frac{1}{1 - \gamma} \\ \begin{align*} v_{\pi}(s_1) &= 0.5[0 + \gamma v_{\pi}(s_3)] + 0.5[-1 + \gamma v_{\pi}(s_2)] \\ & = -0.5 + \frac{\gamma}{1 - \gamma} \end{align*} vπ(s4)=1−γ1vπ(s3)=1−γ1vπ(s2)=1−γ1vπ(s1)=0.5[0+γvπ(s3)]+0.5[−1+γvπ(s2)]=−0.5+1−γγ
Bellman equation: Matrix-vector form
对于Bellman equation:
v π ( s ) = ∑ a π ( a ∣ s ) [ ∑ r p ( r ∣ s , a ) r + γ ∑ s ′ p ( s ′ ∣ s , a ) v π ( s ′ ) ] v_{\pi}(s) = \sum_a \pi(a | s) \left[ \sum_r p(r | s, a)r + \gamma \sum_{s'}p(s' | s, a) v_{\pi}(s') \right] vπ(s)=a∑π(a∣s)[r∑p(r∣s,a)r+γs′∑p(s′∣s,a)vπ(s′)]
通常是未知的 v π ( s ) v_{\pi}(s) vπ(s)伴随着未知的 v π ( s ′ ) v_{\pi}(s') vπ(s′),这对于每一个 s ∈ S s \in \mathcal{S} s∈S都成立。因此,意味着共有 ∣ S ∣ |\mathcal{S}| ∣S∣个这样的等式。如果将所有的等式,放到一起进行计算,这就构成了Bellman equation的矩阵形式。
将上式展开,写为:
v π ( s ) = r π ( s ) + γ ∑ s ′ p π ( s ′ ∣ s ) v π ( s ′ ) ( 1 ) v_{\pi}(s) = r_{\pi}(s) + \gamma \sum_{s'} p_{\pi}(s' | s)v_{\pi}(s') \;\;\;\;\; (1) vπ(s)=rπ(s)+γs′∑pπ(s′∣s)vπ(s′)(1)
其中:
r π ( s ) : = ∑ a π ( a ∣ s ) ∑ r p ( r ∣ s , a ) r p π ( s ′ ∣ s ) : = ∑ a π ( a ∣ s ) p ( s ′ ∣ s , a ) r_{\pi}(s) := \sum_a \pi(a | s) \sum_r p(r | s, a)r \\ p_{\pi}(s' | s) := \sum_a \pi(a | s) p(s' | s, a) rπ(s):=a∑π(a∣s)r∑p(r∣s,a)rpπ(s′∣s):=a∑π(a∣s)p(s′∣s,a)
为state s s s添加索引 s i , i = 1 , . . . , n s_i, i = 1, ..., n si,i=1,...,n
对于 s i s_i si,其Bellman equation为:
v π ( s i ) = r π ( s i ) + γ ∑ s j p π ( s j ∣ s i ) v π ( s j ) v_{\pi}(s_i) = r_{\pi}(s_i) + \gamma \sum_{s_j} p_{\pi}(s_j | s_i)v_{\pi}(s_j) vπ(si)=rπ(si)+γsj∑pπ(sj∣si)vπ(sj)
将所有的state写为矩阵形式:
v π = r π + γ P π v π \mathbf{v}_{\pi} = \mathbf{r}_{\pi} + \gamma \mathbf{P}_{\pi} \mathbf{v}_{\pi} vπ=rπ+γPπvπ
其中:
- v π = [ v π ( s 1 ) , v π ( s 2 ) , . . . , v π ( s n ) ] T ∈ R n \mathbf{v}_{\pi} = [v_{\pi}(s_1), v_{\pi}(s_2), ..., v_{\pi}(s_n)]^T \in \mathbb{R}^n vπ=[vπ(s1),vπ(s2),...,vπ(sn)]T∈Rn
- r π = [ r π ( s 1 ) , r π ( s 2 ) , . . . , r π ( s n ) ] T ∈ R n \mathbf{r}_{\pi} = [r_{\pi}(s_1), r_{\pi}(s_2), ..., r_{\pi}(s_n)]^T \in \mathbb{R}^n rπ=[rπ(s1),rπ(s2),...,rπ(sn)]T∈Rn
- P π ∈ R n × n \mathbf{P}_{\pi} \in \mathbb{R}^{n \times n} Pπ∈Rn×n,其中, [ P π ] = p π ( s j ∣ s i ) [P_{\pi}] = p_{\pi}(s_j | s_i) [Pπ]=pπ(sj∣si)是state转移矩阵。
假设有四个state,则上式矩阵形式可以写为:
[ v π ( s 1 ) v π ( s 2 ) v π ( s 3 ) v π ( s 4 ) ] = [ r π ( s 1 ) r π ( s 2 ) r π ( s 3 ) r π ( s 4 ) ] + γ [ p π ( s 1 ∣ s 1 ) p π ( s 2 ∣ s 1 ) p π ( s 3 ∣ s 1 ) p π ( s 4 ∣ s 1 ) p π ( s 1 ∣ s 2 ) p π ( s 2 ∣ s 2 ) p π ( s 3 ∣ s 2 ) p π ( s 4 ∣ s 2 ) p π ( s 1 ∣ s 3 ) p π ( s 2 ∣ s 3 ) p π ( s 3 ∣ s 3 ) p π ( s 4 ∣ s 3 ) p π ( s 1 ∣ s 4 ) p π ( s 2 ∣ s 4 ) p π ( s 3 ∣ s 4 ) p π ( s 4 ∣ s 4 ) ] [ v π ( s 1 ) v π ( s 2 ) v π ( s 3 ) v π ( s 4 ) ] \begin{bmatrix} v_{\pi}(s_1) \\ v_{\pi}(s_2)\\ v_{\pi}(s_3)\\ v_{\pi}(s_4) \end{bmatrix} = \begin{bmatrix} r_{\pi}(s_1) \\ r_{\pi}(s_2)\\ r_{\pi}(s_3)\\ r_{\pi}(s_4) \end{bmatrix} + \gamma \begin{bmatrix} p_{\pi}(s_1 | s_1) &p_{\pi}(s_2 | s_1) &p_{\pi}(s_3 | s_1) &p_{\pi}(s_4 | s_1)\\ p_{\pi}(s_1 | s_2) &p_{\pi}(s_2 | s_2) &p_{\pi}(s_3 | s_2) &p_{\pi}(s_4 | s_2)\\ p_{\pi}(s_1 | s_3) &p_{\pi}(s_2 | s_3) &p_{\pi}(s_3 | s_3) &p_{\pi}(s_4 | s_3)\\ p_{\pi}(s_1 | s_4) &p_{\pi}(s_2 | s_4) &p_{\pi}(s_3 | s_4) &p_{\pi}(s_4 | s_4) \end{bmatrix} \begin{bmatrix} v_{\pi}(s_1) \\ v_{\pi}(s_2)\\ v_{\pi}(s_3)\\ v_{\pi}(s_4) \end{bmatrix} vπ(s1)vπ(s2)vπ(s3)vπ(s4) = rπ(s1)rπ(s2)rπ(s3)rπ(s4) +γ pπ(s1∣s1)pπ(s1∣s2)pπ(s1∣s3)pπ(s1∣s4)pπ(s2∣s1)pπ(s2∣s2)pπ(s2∣s3)pπ(s2∣s4)pπ(s3∣s1)pπ(s3∣s2)pπ(s3∣s3)pπ(s3∣s4)pπ(s4∣s1)pπ(s4∣s2)pπ(s4∣s3)pπ(s4∣s4) vπ(s1)vπ(s2)vπ(s3)vπ(s4)
例,对policy1:
对其求解,得:
[ v π ( s 1 ) v π ( s 2 ) v π ( s 3 ) v π ( s 4 ) ] = [ 0 1 1 1 ] + γ [ 0 0 1 0 0 0 0 1 0 0 0 1 0 0 0 1 ] [ v π ( s 1 ) v π ( s 2 ) v π ( s 3 ) v π ( s 4 ) ] \begin{bmatrix} v_{\pi}(s_1) \\ v_{\pi}(s_2)\\ v_{\pi}(s_3)\\ v_{\pi}(s_4) \end{bmatrix} = \begin{bmatrix} 0 \\ 1\\ 1\\ 1 \end{bmatrix} + \gamma \begin{bmatrix} 0 &0 &1 &0\\ 0 &0 &0 &1\\ 0 &0 &0 &1\\ 0 &0 &0 &1 \end{bmatrix}\begin{bmatrix} v_{\pi}(s_1) \\ v_{\pi}(s_2)\\ v_{\pi}(s_3)\\ v_{\pi}(s_4) \end{bmatrix} vπ(s1)vπ(s2)vπ(s3)vπ(s4) = 0111 +γ 0000000010000111 vπ(s1)vπ(s2)vπ(s3)vπ(s4)
对policy2:
对其求解,得:
[ v π ( s 1 ) v π ( s 2 ) v π ( s 3 ) v π ( s 4 ) ] = [ 0.5 ( 0 ) + 0.5 ( − 1 ) 1 1 1 ] + γ [ 0 0.5 0.5 0 0 0 0 1 0 0 0 1 0 0 0 1 ] [ v π ( s 1 ) v π ( s 2 ) v π ( s 3 ) v π ( s 4 ) ] \begin{bmatrix} v_{\pi}(s_1) \\ v_{\pi}(s_2)\\ v_{\pi}(s_3)\\ v_{\pi}(s_4) \end{bmatrix} = \begin{bmatrix} 0.5(0) + 0.5(-1) \\ 1\\ 1\\ 1 \end{bmatrix} + \gamma \begin{bmatrix} 0 &0.5 &0.5 &0\\ 0 &0 &0 &1\\ 0 &0 &0 &1\\ 0 &0 &0 &1 \end{bmatrix}\begin{bmatrix} v_{\pi}(s_1) \\ v_{\pi}(s_2)\\ v_{\pi}(s_3)\\ v_{\pi}(s_4) \end{bmatrix} vπ(s1)vπ(s2)vπ(s3)vπ(s4) = 0.5(0)+0.5(−1)111 +γ 00000.50000.50000111 vπ(s1)vπ(s2)vπ(s3)vπ(s4)
Bellman equation: Solve the state values
对于矩阵形式的Bellman equation:
v π = r π + γ P π v π \mathbf{v}_{\pi} = \mathbf{r}_{\pi} + \gamma \mathbf{P}_{\pi} \mathbf{v}_{\pi} vπ=rπ+γPπvπ
其closed-form的解为:
v π = ( I − γ P π ) − 1 r π \mathbf{v}_{\pi} = (\mathbf{I} - \gamma \mathbf{P}_{\pi})^{-1} \mathbf{r}_{\pi} vπ=(I−γPπ)−1rπ
为了避免求矩阵的逆,可以采用迭代法:
v k + 1 = r + γ P π v k v k → v π = ( I − γ P π ) − 1 r π , k → ∞ \mathbf{v}_{k + 1} = \mathbf{r} + \gamma \mathbf{P}_{\pi} \mathbf{v}_k \\ \mathbf{v}_k \rightarrow \mathbf{v}_{\pi} = (\mathbf{I} - \gamma \mathbf{P}_{\pi})^{-1} \mathbf{r}_{\pi}, \;\;\; k \rightarrow \infty vk+1=r+γPπvkvk→vπ=(I−γPπ)−1rπ,k→∞
以下是对于一个grid-world,在给定policy下,各个state的state value。
可以看到,不同的policy其产生的state value可能是相同的。
可以看到,大多数情况下,不同的policy对state value的影响是比较大的,因此,state value是有效评估policy的一个指标。
Action value
state value: agent从某个state开始可以获得的平均return
action value: agent从某个state开始并采取action可以获得的平均return。
通过action value可以知道当前state下,哪个action是更好的。
定义:
q π ( s , a ) = E [ G t ∣ S t = s , A t = a ] q_{\pi}(s, a) = \mathbb{E}[G_t | S_t = s, A_t = a] qπ(s,a)=E[Gt∣St=s,At=a]
- q π ( s , a ) q_{\pi}(s, a) qπ(s,a)是state、action对的函数
- q π ( s , a ) q_{\pi}(s, a) qπ(s,a)依赖于 π \pi π
根据条件期望公式:
E [ G t ∣ S t = s ] = ∑ a E [ G t ∣ S t = s , A t = a ] π ( a ∣ s ) \mathbb{E}[G_t | S_t = s] = \sum_a \mathbb{E}[G_t | S_t = s, A_t = a] \pi (a | s) E[Gt∣St=s]=a∑E[Gt∣St=s,At=a]π(a∣s)
因此,
v π ( s ) = ∑ a π ( a ∣ s ) q π ( s , a ) ( 2 ) v_{\pi}(s) = \sum_{a} \pi(a | s) q_{\pi}(s, a) \;\;\;\;\; (2) vπ(s)=a∑π(a∣s)qπ(s,a)(2)
对于state value:
v π ( s ) = ∑ a π ( a ∣ s ) [ ∑ r p ( r ∣ s , a ) r + γ ∑ s ′ p ( s ′ ∣ s , a ) v π ( s ′ ) ] = ∑ a π ( a ∣ s ) ⋅ q π ( s , a ) ( 3 ) \begin{align*} v_{\pi}(s) &= \sum_a \pi(a | s) \left[ \sum_r p(r | s, a)r + \gamma \sum_{s'}p(s' | s, a) v_{\pi}(s') \right]\\ &=\sum_a \pi(a | s) \cdot q_{\pi}(s, a) \end{align*} \;\;\;\;\; (3) vπ(s)=a∑π(a∣s)[r∑p(r∣s,a)r+γs′∑p(s′∣s,a)vπ(s′)]=a∑π(a∣s)⋅qπ(s,a)(3)
比较公式(2)与公式(3),可以得到action-value function:
q π ( s , a ) = ∑ r p ( r ∣ s , a ) r + γ ∑ s ′ p ( s ′ ∣ s , a ) v π ( s ′ ) ( 4 ) q_{\pi}(s, a) = \sum_r p(r | s, a)r + \gamma \sum_{s'} p(s' | s, a) v_{\pi}(s') \;\;\;\;\; (4) qπ(s,a)=r∑p(r∣s,a)r+γs′∑p(s′∣s,a)vπ(s′)(4)
通过公式(2)和公式(4)可以发现state value和action value可以相互转化。
例:
求解,得:
q π ( s 1 , a 1 ) = − 1 + γ v π ( s 1 ) q π ( s 1 , a 2 ) = − 1 + γ v π ( s 2 ) q π ( s 1 , a 3 ) = 0 + γ v π ( s 3 ) q π ( s 1 , a 4 ) = − 1 + γ v π ( s 1 ) q π ( s 1 , a 5 ) = 0 + γ v π ( s 1 ) \begin{align*} &q_{\pi}(s_1, a_1) = -1 + \gamma v_{\pi}(s_1)\\ &q_{\pi}(s_1, a_2) = -1 + \gamma v_{\pi}(s_2)\\ &q_{\pi}(s_1, a_3) = 0 + \gamma v_{\pi}(s_3) \\ &q_{\pi}(s_1, a_4) = -1 + \gamma v_{\pi}(s_1) \\ &q_{\pi}(s_1, a_5) = 0 + \gamma v_{\pi}(s_1) \end{align*} qπ(s1,a1)=−1+γvπ(s1)qπ(s1,a2)=−1+γvπ(s2)qπ(s1,a3)=0+γvπ(s3)qπ(s1,a4)=−1+γvπ(s1)qπ(s1,a5)=0+γvπ(s1)
Summary
-
state value: v π ( s ) = E [ G t ∣ S t = s ] v_{\pi}(s) = \mathbb{E}[G_t | S_t = s] vπ(s)=E[Gt∣St=s]
-
action value: q π ( s , a ) = E [ G t ∣ S t = s , A t = a ] q_{\pi}(s, a) = \mathbb{E}[G_t | S_t = s, A_t = a] qπ(s,a)=E[Gt∣St=s,At=a]
-
Bellman equation:
elementwise form
v π ( s ) = ∑ a π ( a ∣ s ) [ ∑ r p ( r ∣ s , a ) r + γ ∑ s ′ p ( s ′ ∣ s , a ) v π ( s ′ ) ] = ∑ a π ( a ∣ s ) ⋅ q π ( s , a ) \begin{align*} v_{\pi}(s) &= \sum_a \pi(a | s) \left[ \sum_r p(r | s, a)r + \gamma \sum_{s'}p(s' | s, a) v_{\pi}(s') \right]\\ &=\sum_a \pi(a | s) \cdot q_{\pi}(s, a) \end{align*} vπ(s)=a∑π(a∣s)[r∑p(r∣s,a)r+γs′∑p(s′∣s,a)vπ(s′)]=a∑π(a∣s)⋅qπ(s,a)
matrix-vector form
v π = r π + γ P v π \mathbf{v}_{\pi} = \mathbf{r}_{\pi} + \gamma \mathbf{P} \mathbf{v}_{\pi} vπ=rπ+γPvπ -
可以通过闭合形式解和迭代法求Bellman equation
以上内容为B站西湖大学智能无人系统 强化学习的数学原理 公开课笔记。