考虑下列形式的分布鲁棒优化问题:
min x ∈ X c T x s.t. P [ ξ ~ ∈ S ( x ) ] ≥ 1 − ϵ , ∀ P ∈ F ( θ ) (1) \begin{aligned} \mathop{\min}\limits_{\boldsymbol{x}\in\mathcal{X}} &\quad \boldsymbol{c}^T\boldsymbol{x} \\ \text{s.t.} &\quad \mathbb{P}\left[\tilde{\boldsymbol{\xi}}\in\mathcal{S}(\boldsymbol{x})\right] \ge 1-\epsilon, \; \forall \mathbb{P}\in\mathcal{F}(\theta) \tag{1} \end{aligned} x∈Xmins.t.cTxP[ξ~∈S(x)]≥1−ϵ,∀P∈F(θ)(1) 其中 X ∈ R L \mathcal{X}\in\mathbb{R}^L X∈RL 是一个紧多面体, S ( x ) ⊆ R K \mathcal{S}(\boldsymbol{x})\subseteq\mathbb{R}^K S(x)⊆RK 是一个决策相关安全集( decision-dependent safety set) , F ( θ ) \mathcal{F}(\theta) F(θ) 是以 P ^ \hat{\mathbb{P}} P^ 为中心、 θ \theta θ 为半径的 Wasserstein Ball。这里 P ^ \hat{\mathbb{P}} P^ 是一组训练数据 { ξ ^ i } i ∈ [ N ] \left\{\hat{\boldsymbol{\xi}}_i\right\}_{i\in[N]} {ξ^i}i∈[N] 的经验分布。
假设 S ⊆ R K \mathcal{S}\subseteq\mathbb{R}^K S⊆RK,记 S ˉ = R K ∖ S \bar{\mathcal{S}}=\mathbb{R}^K\setminus\mathcal{S} Sˉ=RK∖S 为它的闭补集。考虑下面的不确定度量问题:
sup P ∈ F ( θ ) P [ ξ ~ ∉ S ] . \mathop{\sup}\limits_{\mathbb{P}\in\mathcal{F}(\theta)} \mathbb{P}\left[\tilde{\boldsymbol{\xi}}\notin\mathcal{S}\right]. P∈F(θ)supP[ξ~∈/S]. 这里都假设 θ > 0 , ϵ ∈ ( 0 , 1 ) \theta>0,\epsilon\in(0,1) θ>0,ϵ∈(0,1)。
记号:
(1) 记 dist ( ξ ^ i , S ˉ ) \text{dist}\left(\hat{\boldsymbol{\xi}}_i, \bar{\mathcal{S}}\right) dist(ξ^i,Sˉ) 为 P ^ \hat{\mathbb{P}} P^ 对应的第 i i i 个数据点和不安全集 S ˉ \bar{\mathcal{S}} Sˉ 之间的距离,这里的距离是基于某个范数 ∥ ⋅ ∥ \lVert\cdot\rVert ∥⋅∥。我们对样本数据集进行排序,使其满足
dist ( ξ ^ i , S ˉ ) ≤ dist ( ξ ^ j , S ˉ ) , ∀ 1 ≤ i ≤ j ≤ N \text{dist}\left(\hat{\boldsymbol{\xi}}_i, \bar{\mathcal{S}}\right) \leq \text{dist}\left(\hat{\boldsymbol{\xi}}_j, \bar{\mathcal{S}}\right), \quad \forall 1\leq i\leq j\leq N dist(ξ^i,Sˉ)≤dist(ξ^j,Sˉ),∀1≤i≤j≤N (2) 假设 dist ( ξ ^ i , S ˉ ) = 0 \text{dist}\left(\hat{\boldsymbol{\xi}}_i, \bar{\mathcal{S}}\right) =0 dist(ξ^i,Sˉ)=0 当且仅当 i ∈ [ I ] i\in[I] i∈[I], 这里 I I I 可以为0。
(3) 记 ξ i ∗ ∈ S ˉ \boldsymbol{\xi}_i^*\in\bar{\mathcal{S}} ξi∗∈Sˉ 为距离 ξ ^ i \hat{\boldsymbol{\xi}}_i ξ^i 最近的不安全点。
定理. (Blanchet and Murthy 2016, Gao and Kleywegt 2016)
令 j ∗ = max { j ∈ [ N ] ∪ { 0 } ∣ ∑ i = 1 j dist ( ξ ^ i , S ˉ ) ≤ θ N } j^*=\max\left\{ j\in[N]\cup\{0\} \;\big|\; \mathop{\sum}\limits_{i=1}^j \text{dist}\left(\hat{\boldsymbol{\xi}}_i, \bar{\mathcal{S}}\right) \leq \theta N \right\} j∗=max{j∈[N]∪{0}∣∣i=1∑jdist(ξ^i,Sˉ)≤θN}. 则上述不确定度量问题可以由下面给出的最差情形分布 P ∗ ∈ F ( θ ) \mathbb{P}^*\in\mathcal{F}(\theta) P∗∈F(θ) 来求解:
(1) 如果 j ∗ = N j^*=N j∗=N,则 sup P ∈ F ( θ ) P [ ξ ~ ∉ S ] = P ∗ [ ξ ~ ∉ S ] = 1 \mathop{\sup}\limits_{\mathbb{P}\in\mathcal{F}(\theta)} \mathbb{P}\left[\tilde{\boldsymbol{\xi}}\notin\mathcal{S}\right] = \mathbb{P}^*\left[\tilde{\boldsymbol{\xi}}\notin\mathcal{S}\right]=1 P∈F(θ)supP[ξ~∈/S]=P∗[ξ~∈/S]=1,其中
P ∗ = 1 N ∑ i = 1 I δ ξ ^ i + 1 N ∑ i = I + 1 N δ ξ i ∗ \mathbb{P}^* = \frac{1}{N}\mathop{\sum}\limits_{i=1}^I\delta_{\hat{\boldsymbol{\xi}}_i} + \frac{1}{N}\mathop{\sum}\limits_{i=I+1}^N\delta_{\boldsymbol{\xi}^*_i} P∗=N1i=1∑Iδξ^i+N1i=I+1∑Nδξi∗
(2) 如果 j ∗ < N j^*
对任何给定的决策变量 x ∈ X \boldsymbol{x}\in\mathcal{X} x∈X,存在 [ N ] [N] [N]的一个置换 π ( x ) \pi(\boldsymbol{x}) π(x),使得
dist ( ξ ^ π 1 ( x ) , S ˉ ( x ) ) ≤ dist ( ξ ^ π 2 ( x ) , S ˉ ( x ) ) ≤ ⋯ dist ( ξ ^ π N ( x ) , S ˉ ( x ) ) \text{dist}\left(\hat{\boldsymbol{\xi}}_{\pi_1(\boldsymbol{x})}, \bar{\mathcal{S}}(\boldsymbol{x})\right) \leq \text{dist}\left(\hat{\boldsymbol{\xi}}_{\pi_2(\boldsymbol{x})}, \bar{\mathcal{S}}(\boldsymbol{x})\right) \leq \cdots \text{dist}\left(\hat{\boldsymbol{\xi}}_{\pi_N(\boldsymbol{x})}, \bar{\mathcal{S}}(\boldsymbol{x})\right) dist(ξ^π1(x),Sˉ(x))≤dist(ξ^π2(x),Sˉ(x))≤⋯dist(ξ^πN(x),Sˉ(x))
定理. 对任何给定的决策变量 x ∈ X \boldsymbol{x}\in\mathcal{X} x∈X,模糊机会约束问题 (1) 等价于下面的确定性不等式: 1 N ∑ i = 1 ϵ N dist ( ξ ^ π i ( x ) , S ˉ ( x ) ) ≥ θ . \frac{1}{N}\mathop{\sum}\limits_{i=1}^{\epsilon N} \text{dist}\left(\hat{\boldsymbol{\xi}}_{\pi_i(\boldsymbol{x})}, \bar{\mathcal{S}}(\boldsymbol{x})\right) \ge \theta. N1i=1∑ϵNdist(ξ^πi(x),Sˉ(x))≥θ.
进一步,利用线性对偶,可以证明:
定理. 模糊机会约束问题 (1) 等价于:
min s , t , x c T x s.t. ϵ N t − e T s ≥ θ N dist ( ξ ^ i , S ˉ ( x ) ) ≥ t − s i , ∀ i ∈ [ N ] s ≥ 0 , x ∈ X . \begin{aligned} \mathop{\min}\limits_{\boldsymbol{s},t,\boldsymbol{x}} &\quad \boldsymbol{c}^T\boldsymbol{x} \\ \text{s.t.} &\quad \epsilon Nt - \boldsymbol{e}^T\boldsymbol{s} \ge \theta N \\ &\quad \text{dist}\left(\hat{\boldsymbol{\xi}}_i, \bar{\mathcal{S}}(\boldsymbol{x})\right) \ge t-s_i, \quad \forall i\in[N] \\ &\quad \boldsymbol{s}\ge\boldsymbol{0},\boldsymbol{x}\in\mathcal{X}. \end{aligned} s,t,xmins.t.cTxϵNt−eTs≥θNdist(ξ^i,Sˉ(x))≥t−si,∀i∈[N]s≥0,x∈X.
命题. 假设对所有 x ∈ X \boldsymbol{x}\in\mathcal{X} x∈X, A T x ≠ b \boldsymbol{A}^T\boldsymbol{x}\neq\boldsymbol{b} ATx=b。对于安全集 S ( x ) = { ξ ∈ R K ∣ ( A ξ + a ) T x < b T ξ + b } \mathcal{S}(\boldsymbol{x}) = \{ \boldsymbol{\xi}\in\mathbb{R}^K \;|\; (\boldsymbol{A\xi+a})^T\boldsymbol{x} < \boldsymbol{b}^T\boldsymbol{\xi} + b \} S(x)={ξ∈RK∣(Aξ+a)Tx<bTξ+b}。问题 (1) 等价于:
Z I C C ∗ = min q , s , t , x c T x s.t. ϵ N t − e T s ≥ θ N ∥ b − A T x ∥ ∗ ( b − A T x ) T ξ ^ i + b − a T x + M q i ≥ t − s i , ∀ i ∈ [ N ] M ( 1 − q i ) ≥ t − s i , ∀ i ∈ [ N ] q ∈ { 0 , 1 } N , s ≥ 0 , x ∈ X . \begin{aligned} Z_{ICC}^*=\mathop{\min}\limits_{\boldsymbol{q},\boldsymbol{s},t,\boldsymbol{x}} &\quad \boldsymbol{c}^T\boldsymbol{x} \\ \text{s.t.} &\quad \epsilon Nt - \boldsymbol{e}^T\boldsymbol{s} \ge \theta N \lVert \boldsymbol{b}-\boldsymbol{A}^T\boldsymbol{x}\rVert_* \\ &\quad ( \boldsymbol{b}-\boldsymbol{A}^T\boldsymbol{x})^T\hat{\boldsymbol{\xi}}_i+b-\boldsymbol{a}^T\boldsymbol{x}+Mq_i \ge t-s_i, \quad \forall i\in[N] \\ &\quad M(1-q_i)\ge t-s_i , \quad \forall i\in[N] \\ &\quad \boldsymbol{q}\in\{0,1\}^N,\boldsymbol{s}\ge\boldsymbol{0},\boldsymbol{x}\in\mathcal{X}. \end{aligned} ZICC∗=q,s,t,xmins.t.cTxϵNt−eTs≥θN∥b−ATx∥∗(b−ATx)Tξ^i+b−aTx+Mqi≥t−si,∀i∈[N]M(1−qi)≥t−si,∀i∈[N]q∈{0,1}N,s≥0,x∈X.
命题. 对于安全集 S ( x ) = { ξ ∈ R K ∣ a m T x < b m T ξ + b m } \mathcal{S}(\boldsymbol{x}) = \{ \boldsymbol{\xi}\in\mathbb{R}^K \;|\; \boldsymbol{a}_m^T\boldsymbol{x} < \boldsymbol{b}_m^T\boldsymbol{\xi} + b_m \} S(x)={ξ∈RK∣amTx<bmTξ+bm}。问题 (1) 等价于:
Z J C C ∗ = min p , q , s , t , x c T x s.t. ϵ N t − e T s ≥ θ N p i + M q i ≥ t − s i , ∀ i ∈ [ N ] M ( 1 − q i ) ≥ t − s i , ∀ i ∈ [ N ] b m T ξ ^ i + b m − a m T x ∥ b m ∥ ∗ ≥ p i , ∀ m ∈ [ M ] , ∀ i ∈ [ N ] q ∈ { 0 , 1 } N , s ≥ 0 , x ∈ X . \begin{aligned} Z_{JCC}^*=\mathop{\min}\limits_{\boldsymbol{p},\boldsymbol{q},\boldsymbol{s},t,\boldsymbol{x}} &\quad \boldsymbol{c}^T\boldsymbol{x} \\ \text{s.t.} &\quad \epsilon Nt - \boldsymbol{e}^T\boldsymbol{s} \ge \theta N \\ &\quad p_i+Mq_i\ge t-s_i, \quad \forall i\in[N] \\ &\quad M(1-q_i)\ge t-s_i, \quad \forall i\in[N] \\ &\quad \frac{\boldsymbol{b}^T_m\hat{\boldsymbol{\xi}}_i+b_m-\boldsymbol{a}^T_m\boldsymbol{x}}{\lVert\boldsymbol{b}_m\rVert_*} \ge p_i, \quad \forall m\in[M],\forall i\in[N] \\ &\quad \boldsymbol{q}\in\{0,1\}^N,\boldsymbol{s}\ge\boldsymbol{0},\boldsymbol{x}\in\mathcal{X}. \end{aligned} ZJCC∗=p,q,s,t,xmins.t.cTxϵNt−eTs≥θNpi+Mqi≥t−si,∀i∈[N]M(1−qi)≥t−si,∀i∈[N]∥bm∥∗bmTξ^i+bm−amTx≥pi,∀m∈[M],∀i∈[N]q∈{0,1}N,s≥0,x∈X.
Z I C C ∗ = min ( x , s , t ) ∈ C I C C c T x Z_{ICC}^*=\mathop{\min}\limits_{(\boldsymbol{x},\boldsymbol{s},t)\in\mathcal{C}_{ICC} } \boldsymbol{c}^T\boldsymbol{x} ZICC∗=(x,s,t)∈CICCmincTx
其中
C I C C = { ( x , s , t ) ∈ X × R + N × R ∣ ϵ N t − e T s ≥ θ N ∥ b − A T x ∥ ∗ ( ( b − A T x ) T ξ ^ i + b − a T x ) + ≥ t − s i , ∀ i ∈ [ N ] } \mathcal{C}_{ICC} =\left\{ (\boldsymbol{x},\boldsymbol{s},t)\in\mathcal{X}\times\mathbb{R}^N_+\times\mathbb{R} \;\Bigg|\; \begin{aligned} &\; \epsilon Nt - \boldsymbol{e}^T\boldsymbol{s} \ge \theta N \lVert \boldsymbol{b}-\boldsymbol{A}^T\boldsymbol{x}\rVert_* \\ &\; \left( ( \boldsymbol{b}- \boldsymbol{A}^T\boldsymbol{x})^T\hat{\boldsymbol{\xi}}_i+b-\boldsymbol{a}^T\boldsymbol{x} \right)^+ \ge t-s_i, \quad \forall i\in[N] \end{aligned} \right\} CICC=⎩⎨⎧(x,s,t)∈X×R+N×R∣∣∣∣∣ϵNt−eTs≥θN∥b−ATx∥∗((b−ATx)Tξ^i+b−aTx)+≥t−si,∀i∈[N]⎭⎬⎫
C I C C \mathcal{C}_{ICC} CICC 是非-凸的,可以需要找一些凸集来逼近它。
C I C C ( k ) = { ( x , s , t ) ∈ X × R + N × R ∣ ϵ N t − e T s ≥ θ N ∥ b − A T x ∥ ∗ k i ( ( b − A T x ) T ξ ^ i + b − a T x ) ≥ t − s i , ∀ i ∈ [ N ] } \mathcal{C}_{ICC}(\boldsymbol{k}) =\left\{ (\boldsymbol{x},\boldsymbol{s},t)\in\mathcal{X}\times\mathbb{R}^N_+\times\mathbb{R} \;\Bigg|\; \begin{aligned} &\; \epsilon Nt - \boldsymbol{e}^T\boldsymbol{s} \ge \theta N \lVert \boldsymbol{b}-\boldsymbol{A}^T\boldsymbol{x}\rVert_* \\ &\; k_i \left(( \boldsymbol{b}- \boldsymbol{A}^T\boldsymbol{x})^T\hat{\boldsymbol{\xi}}_i+b-\boldsymbol{a}^T\boldsymbol{x} \right) \ge t-s_i, \quad \forall i\in[N] \end{aligned} \right\} CICC(k)=⎩⎨⎧(x,s,t)∈X×R+N×R∣∣∣∣∣ϵNt−eTs≥θN∥b−ATx∥∗ki((b−ATx)Tξ^i+b−aTx)≥t−si,∀i∈[N]⎭⎬⎫
命题. 对于任何凸集 W ⊆ C I C C \mathcal{W}\subseteq\mathcal{C}_{ICC} W⊆CICC,存在 k ∈ [ 0 , 1 ] N \boldsymbol{k}\in[0,1]^N k∈[0,1]N,使得 W ⊆ C I C C ( k ) ⊆ C I C C \mathcal{W}\subseteq\mathcal{C}_{ICC}(\boldsymbol{k}) \subseteq\mathcal{C}_{ICC} W⊆CICC(k)⊆CICC。
Z I C C ∗ ( k ) = min ( x , s , t ) ∈ C I C C ( k ) c T x Z_{ICC}^*(\boldsymbol{k})=\mathop{\min}\limits_{(\boldsymbol{x},\boldsymbol{s},t)\in\mathcal{C}_{ICC}(\boldsymbol{k}) } \boldsymbol{c}^T\boldsymbol{x} ZICC∗(k)=(x,s,t)∈CICC(k)mincTx
命题. Z I C C ∗ = min k ∈ [ 0 , 1 ] N Z I C C ∗ ( k ) Z_{ICC}^*= \mathop{\min}\limits_{\boldsymbol{k}\in[0,1]^N} Z_{ICC}^*(\boldsymbol{k}) ZICC∗=k∈[0,1]NminZICC∗(k)。
命题. min k ∈ [ 0 , 1 ] Z I C C ∗ ( k e ) = Z I C C ∗ ( e ) \mathop{\min}\limits_{k\in[0,1]} Z_{ICC}^*(k\boldsymbol{e})= Z_{ICC}^*(\boldsymbol{e}) k∈[0,1]minZICC∗(ke)=ZICC∗(e)。
P [ ξ ~ ∈ S ( x ) ] ≥ 1 − ϵ ⟺ P [ ( A ξ ~ + a ) T x ≥ b T ξ ~ + b ] ≤ ϵ ⟺ P -VaR ϵ ( a T x − b + ( A T x − b ) T ξ ~ ) ≤ 0 ⟸ P -CVaR ϵ ( a T x − b + ( A T x − b ) T ξ ~ ) ≤ 0 \begin{aligned} &\quad \mathbb{P}[\tilde{\boldsymbol{\xi}} \in \mathcal{S}(\boldsymbol{x})] \ge 1-\epsilon \\ \Longleftrightarrow &\quad \mathbb{P}\left[ (\boldsymbol{A}\tilde{\boldsymbol{\xi}} + \boldsymbol{a} )^T \boldsymbol{x} \ge \boldsymbol{b}^T\tilde{\boldsymbol{\xi}}+b \right] \leq \epsilon \\ \Longleftrightarrow &\quad \mathbb{P}\text{-VaR} _{\epsilon} \left( \boldsymbol{a} ^T \boldsymbol{x} - b + (\boldsymbol{A}^T\boldsymbol{x}-\boldsymbol{b})^T\tilde{\boldsymbol{\xi}} \right) \leq 0 \\ \textcolor{red}{\Longleftarrow} &\quad \mathbb{P}\text{-CVaR} _{\epsilon} \left( \boldsymbol{a} ^T \boldsymbol{x} - b + (\boldsymbol{A}^T\boldsymbol{x}-\boldsymbol{b})^T\tilde{\boldsymbol{\xi}} \right) \leq 0 \end{aligned} ⟺⟺⟸P[ξ~∈S(x)]≥1−ϵP[(Aξ~+a)Tx≥bTξ~+b]≤ϵP-VaRϵ(aTx−b+(ATx−b)Tξ~)≤0P-CVaRϵ(aTx−b+(ATx−b)Tξ~)≤0
Z CVaR ∗ = { min x ∈ X c T x s.t. sup P ∈ F ( θ ) P -CVaR ϵ ( a T x − b + ( A T x − b ) T ξ ~ ) ≤ 0 \begin{aligned} Z_{\text{CVaR}}^* = \begin{cases} \mathop{\min}\limits_{\boldsymbol{x}\in\mathcal{X}} \; \boldsymbol{c}^T\boldsymbol{x} \\ \text{s.t.} \; \mathop{\sup}\limits_{\mathbb{P}\in\mathcal{F}(\theta)} \mathbb{P}\text{-CVaR} _{\epsilon} \left( \boldsymbol{a} ^T \boldsymbol{x} - b + (\boldsymbol{A}^T\boldsymbol{x}-\boldsymbol{b})^T\tilde{\boldsymbol{\xi}} \right) \leq 0 \end{cases} \end{aligned} ZCVaR∗=⎩⎪⎨⎪⎧x∈XmincTxs.t.P∈F(θ)supP-CVaRϵ(aTx−b+(ATx−b)Tξ~)≤0
命题. Z CVaR ∗ = Z ICC ∗ ( e ) Z_{\text{CVaR}}^* = Z_{\text{ICC}}^* (\boldsymbol{e}) ZCVaR∗=ZICC∗(e)