优化 | 线性化:含绝对值的线性化
作者:刘兴禄, 清华大学
清华-伯克利深圳学院,博士在读
欢迎关注我们的微信公众号 运小筹
非线性整数规划模型
考虑下面的非线性整数规划
max 2 ∣ x 1 ∣ + ∣ x 2 ∣ s . t . ∣ x 1 ∣ + 2 ∣ x 2 ∣ ⩽ 8 − 5 ⩽ x 1 , x 2 ⩽ 5 \begin{aligned} \max \quad \,\,\,& 2|x_1| + |x_2| \\ s.t. \quad \,\,\,& |x_1| + 2|x_2| \leqslant 8 \\ & -5 \leqslant x_1, x_2 \leqslant 5 \end{aligned} maxs.t.2∣x1∣+∣x2∣∣x1∣+2∣x2∣⩽8−5⩽x1,x2⩽5
Gurobi求解代码
from gurobipy import *
model = Model('non-linear model')
x1 = model.addVar(lb=-5, ub=5, vtype=GRB.CONTINUOUS, name='x1')
x2 = model.addVar(lb=-5, ub=5, vtype=GRB.CONTINUOUS, name='x2')
x1_abs = model.addVar(lb=0, ub=5, vtype=GRB.CONTINUOUS, name='x1_abs')
x2_abs = model.addVar(lb=0, ub=5, vtype=GRB.CONTINUOUS, name='x2_abs')
model.setObjective(2 * x1_abs + x2_abs, GRB.MAXIMIZE)
model.addConstr(x1_abs + 2*x2_abs <= 8)
model.addGenConstrAbs(x1_abs, x1)
model.addGenConstrAbs(x2_abs, x2)
model.optimize()
print('x1:', x1.x)
print('x2:', x2.x)
print('x1_abs:', x1_abs.x)
print('x2_abs:', x2_abs.x)
求解结果
Solution count 1: 11.5
Optimal solution found (tolerance 1.00e-04)
Best objective 1.150000000000e+01, best bound 1.150000000000e+01, gap 0.0000%
x1: 5.0
x2: 1.5
x1_abs: 5.0
x2_abs: 1.5
绝对值的线性化技巧
我们引入两组辅助变量 x i + , x i − , ∀ i = 1 , 2 x_i^{+}, x_i^{-}, \forall i =1,2 xi+,xi−,∀i=1,2。其中
x i + = max { 0 , x i } x i − = max { 0 , − x i } \begin{aligned} &x_i^{+} = \max\{0, x_i\} \\ &x_i^{-} = \max\{0, -x_i\} \end{aligned} xi+=max{0,xi}xi−=max{0,−xi}
因此,就有
x i = x i + − x i − ∣ x i ∣ = x i + + x i − \begin{aligned} &x_i = x_i^{+} - x_i^{-} \\ &|x_i| = x_i^{+} + x_i^{-} \end{aligned} xi=xi+−xi−∣xi∣=xi++xi−
这个比较容易理解,例如
x
i
=
−
0.5
x_i = -0.5
xi=−0.5, 则
x
i
+
=
0
,
x
i
−
=
0.5
x_i^{+} = 0, x_i^{-} = 0.5
xi+=0,xi−=0.5, 因此有
x
i
=
x
i
+
−
x
i
−
=
0
−
0.5
=
−
0.5
x_i = x_i^{+} - x_i^{-} = 0 - 0.5 = -0.5
xi=xi+−xi−=0−0.5=−0.5;
∣
x
i
∣
=
x
i
+
+
x
i
−
=
0
+
0.5
=
0.5
|x_i| = x_i^{+} + x_i^{-} = 0 + 0.5 = 0.5
∣xi∣=xi++xi−=0+0.5=0.5。
拓展
之前的推文中,有小伙伴纠结这个转化是否等价。一个肯定的回答是:完全等价。
这里我给出证明:
命题
:
x + = max { 0 , x } ( 1 ) x − = max { 0 , − x } ( 2 ) \begin{aligned} &x^+=\max \left\{ 0,x \right\} \,\, \left( 1 \right) \\ &x^-=\max \left\{ 0,-x \right\} \,\, \left( 2 \right) \end{aligned} x+=max{0,x}(1)x−=max{0,−x}(2)
和
x = x + − x − ( 3 ) ∣ x ∣ = x + + x − ( 4 ) \begin{aligned} &x=x^+-x^-\,\, \left( 3 \right) \\ &|x|=x^++x^-\,\,\left( 4 \right) \end{aligned} x=x+−x−(3)∣x∣=x++x−(4)
互为充要条件。证明:
充分性:根据(1)(2),我们有
x = x + − x − ( 3 ) ∣ x ∣ = x + + x − ( 4 ) \begin{aligned} &x=x^+-x^-\,\, \left( 3 \right) \\ &|x|=x^++x^-\,\,\left( 4 \right) \end{aligned} x=x+−x−(3)∣x∣=x++x−(4)
充分性得证
必要性: 根据(3)(4)我们有
( 3 ) + ( 4 ) = x + ∣ x ∣ = 2 x + → x + = x + ∣ x ∣ 2 ( 3 ) − ( 4 ) = x − ∣ x ∣ = − 2 x − → x − = − x − ∣ x ∣ 2 情况一: i f x ⩾ 0 , x + = x + ∣ x ∣ 2 = x + x 2 = x = max { 0 , x } x − = − x − ∣ x ∣ 2 = − x − x 2 = 0 = max { 0 , − x } 情况二: i f x ⩽ 0 , x + = x + ∣ x ∣ 2 = x − x 2 = 0 = max { 0 , x } x − = − x − ∣ x ∣ 2 = − x − ( − x ) 2 = − 2 x 2 = − x = max { 0 , − x } \begin{aligned} &\left( 3 \right) +\left( 4 \right) =x+|x|=2x^+\,\, \rightarrow \,\,x^+=\frac{x+|x|}{2} \\ &\left( 3 \right) -\left( 4 \right) =x-|x|=-2x^-\,\, \rightarrow \,\,x^-=-\frac{x-|x|}{2} \\ &\text{情况一:}if\,\,x\geqslant 0, x^+=\frac{x+|x|}{2}=\frac{x+x}{2}=x=\max \left\{ 0,x \right\} \\ &x^-=-\frac{x-|x|}{2}=-\frac{x-x}{2}=0=\max \left\{ 0,-x \right\} \\ &\text{情况二:}if\,\,x\leqslant 0, x^+=\frac{x+|x|}{2}=\frac{x-x}{2}=0=\max \left\{ 0,x \right\} \\ &x^-=-\frac{x-|x|}{2}=-\frac{x-\left( -x \right)}{2}=-\frac{2x}{2}=-x=\max \left\{ 0,-x \right\} \end{aligned} (3)+(4)=x+∣x∣=2x+→x+=2x+∣x∣(3)−(4)=x−∣x∣=−2x−→x−=−2x−∣x∣情况一:ifx⩾0,x+=2x+∣x∣=2x+x=x=max{0,x}x−=−2x−∣x∣=−2x−x=0=max{0,−x}情况二:ifx⩽0,x+=2x+∣x∣=2x−x=0=max{0,x}x−=−2x−∣x∣=−2x−(−x)=−22x=−x=max{0,−x}
因此,必要性得证。
综上,(1)(2)和(3)(4)互为充要条件,以上转化完全等价。
利用上面的技巧进行线性化
- 我们根据
我们引入两组辅助变量 x i + , x i − , ∀ i = 1 , 2 x_i^{+}, x_i^{-}, \forall i =1,2 xi+,xi−,∀i=1,2。其中
x i + = max { 0 , x i } x i − = max { 0 , − x i } \begin{aligned} &x_i^{+} = \max\{0, x_i\} \\ &x_i^{-} = \max\{0, -x_i\} \end{aligned} xi+=max{0,xi}xi−=max{0,−xi}
因此,就有
x i = x i + − x i − ∣ x i ∣ = x i + + x i − \begin{aligned} &x_i = x_i^{+} - x_i^{-} \\ &|x_i| = x_i^{+} + x_i^{-} \end{aligned} xi=xi+−xi−∣xi∣=xi++xi−
- 由于目标函数为 max \max max,因此还需要引入 x 1 abs x_1^{\text{abs}} x1abs和 x 2 abs x_2^{\text{abs}} x2abs来辅助,否则会出现:
- x 1 = 0 , x 1 + = 5 , x 1 − = 5 x_1 = 0, x_1^{+} =5, x_1^{-}=5 x1=0,x1+=5,x1−=5, 也满足 x 1 = x 1 + − x 1 − x_1 = x_1^{+} - x_1^{-} x1=x1+−x1−
- 因此需要把 ∣ x i ∣ = x i + + x i − |x_i| = x_i^{+} + x_i^{-} ∣xi∣=xi++xi−的部分也引入。
如果目标函数是 min ∣ x ∣ \min |x| min∣x∣,我们可以引入辅助变量 y y y,并且加入约束
y ⩾ x y ⩾ − x \begin{aligned} &y \geqslant x \\ &y \geqslant -x \end{aligned} y⩾xy⩾−x
然后把目标改为(也就是完整形式)
min y s . t . y ⩾ x y ⩾ − x \begin{aligned} \min \,\,\, y \\ s.t. \,\,\, &y \geqslant x \\ &y \geqslant -x \end{aligned} minys.t.y⩾xy⩾−x
即可。 但是如果目标是 max \max max的形式,上述情况不再适用。就可以用本文的方法。
得到,原模型可以线性化为
max 2 ( x 1 + + x 1 − ) + ( x 2 + + x 2 − ) s . t . ( x 1 + + x 1 − ) + 2 ( x 2 + + x 2 − ) ⩽ 8 x 1 = x 1 + − x 1 − x 2 = x 2 + − x 2 − x 1 abs = x 1 + + x 1 − x 2 abs = x 2 + + x 2 − x 1 abs ⩾ x 1 x 1 abs ⩾ − x 1 x 2 abs ⩾ x 2 x 2 abs ⩾ − x 2 − 5 ⩽ x 1 , x 2 ⩽ 5 0 ⩽ x i + , x i − , x i abs ⩽ 5 , ∀ i = 1 , 2 \begin{aligned} \max \quad \,\,\,& 2(x_1^{+} + x_1^{-}) + (x_2^{+} + x_2^{-}) \\ s.t. \quad \,\,\,& (x_1^{+} + x_1^{-}) + 2(x_2^{+} + x_2^{-}) \leqslant 8 \\ & x_1 = x_1^{+} - x_1^{-} \\ & x_2 = x_2^{+} - x_2^{-} \\ & x_1^{\text{abs}} = x_1^{+} + x_1^{-} \\ & x_2^{\text{abs}} = x_2^{+} + x_2^{-} \\ & x_1^{\text{abs}} \geqslant x_1 \\ & x_1^{\text{abs}} \geqslant -x_1 \\ & x_2^{\text{abs}} \geqslant x_2 \\ & x_2^{\text{abs}} \geqslant -x_2 \\ & -5 \leqslant x_1, x_2 \leqslant 5 \\ & 0 \leqslant x_i^{+}, x_i^{-}, x_i^{\text{abs}}\leqslant 5, \forall i = 1, 2 \end{aligned} maxs.t.2(x1++x1−)+(x2++x2−)(x1++x1−)+2(x2++x2−)⩽8x1=x1+−x1−x2=x2+−x2−x1abs=x1++x1−x2abs=x2++x2−x1abs⩾x1x1abs⩾−x1x2abs⩾x2x2abs⩾−x2−5⩽x1,x2⩽50⩽xi+,xi−,xiabs⩽5,∀i=1,2
下面我们用代码进行验证。
from gurobipy import *
model = Model('non-linear model')
x1 = model.addVar(lb=-5, ub=5, vtype=GRB.CONTINUOUS, name='x1')
x2 = model.addVar(lb=-5, ub=5, vtype=GRB.CONTINUOUS, name='x2')
x1_pos_abs = model.addVar(lb=0, ub=5, vtype=GRB.CONTINUOUS, name='x1_pos_abs')
x1_neg_abs = model.addVar(lb=0, ub=5, vtype=GRB.CONTINUOUS, name='x1_neg_abs')
x2_pos_abs = model.addVar(lb=0, ub=5, vtype=GRB.CONTINUOUS, name='x2_pos_abs')
x2_neg_abs = model.addVar(lb=0, ub=5, vtype=GRB.CONTINUOUS, name='x2_neg_abs')
x1_abs = model.addVar(lb=0, ub=5, vtype=GRB.CONTINUOUS, name='x1_abs')
x2_abs = model.addVar(lb=0, ub=5, vtype=GRB.CONTINUOUS, name='x2_abs')
model.setObjective(2*(x1_pos_abs + x1_neg_abs) + (x2_pos_abs + x2_neg_abs), GRB.MAXIMIZE)
model.addConstr((x1_pos_abs + x1_neg_abs) + 2*(x2_pos_abs + x2_neg_abs) <= 8)
model.addConstr(x1 == x1_pos_abs - x1_neg_abs)
model.addConstr(x2 == x2_pos_abs - x2_neg_abs)
model.addConstr(x1_abs == x1_pos_abs + x1_neg_abs)
model.addConstr(x2_abs == x2_pos_abs + x2_neg_abs)
model.addConstr(x1_abs >= x1)
model.addConstr(x1_abs >= -x1)
model.addConstr(x2_abs >= x2)
model.addConstr(x2_abs >= -x2)
model.optimize()
print('x1:', x1.x)
print('x2:', x2.x)
print('x1_pos_abs:', x1_pos_abs.x)
print('x1_neg_abs:', x1_neg_abs.x)
print('x2_pos_abs:', x2_pos_abs.x)
print('x2_neg_abs:', x2_neg_abs.x)
求解结果为
Iteration Objective Primal Inf. Dual Inf. Time
0 1.6000000e+01 3.000000e+00 0.000000e+00 0s
2 1.1500000e+01 0.000000e+00 0.000000e+00 0s
Solved in 2 iterations and 0.01 seconds
Optimal objective 1.150000000e+01
x1: 5.0
x2: 1.5
x1_pos_abs: 5.0
x1_neg_abs: 0.0
x2_pos_abs: 1.5
x2_neg_abs: 0.0
可见结果是一致的。
注意,如果约束里有一个不是绝对值,会出错的。经过了测试 x 1 + 2 ∣ x 2 ∣ ⩽ 8 x_1 + 2|x_2| \leqslant 8 x1+2∣x2∣⩽8,上述引入了 x 1 abs x_1^{\text{abs}} x1abs的方法也是等价的。
- 但是,下面的情形是不等价的,如果是 max 2 ∣ x 1 ∣ + ∣ x 2 ∣ s . t . x 1 + 2 ∣ x 2 ∣ ⩽ 8 − 5 ⩽ x 1 , x 2 ⩽ 5 \begin{aligned} \max \quad \,\,\,& 2|x_1| + |x_2| \\ s.t. \quad \,\,\,& x_1 + 2|x_2| \leqslant 8 \\ & -5 \leqslant x_1, x_2 \leqslant 5 \end{aligned} maxs.t.2∣x1∣+∣x2∣x1+2∣x2∣⩽8−5⩽x1,x2⩽5
然后我们转化成
max 2 ( x 1 + + x 1 − ) + ( x 2 + + x 2 − ) s . t . x 1 + 2 ( x 2 + + x 2 − ) ⩽ 8 x 1 = x 1 + − x 1 − x 2 = x 2 + − x 2 − ( 可以删去 ) − 5 ⩽ x 1 , x 2 ⩽ 5 0 ⩽ x i + , x i − ⩽ 5 , ∀ i = 1 , 2 \begin{aligned} \max \quad \,\,\,& 2(x_1^{+} + x_1^{-}) + (x_2^{+} + x_2^{-}) \\ s.t. \quad \,\,\,& x_1 + 2(x_2^{+} + x_2^{-}) \leqslant 8 \\ & x_1 = x_1^{+} - x_1^{-} \\ & x_2 = x_2^{+} - x_2^{-} (\text{可以删去}) \\ & -5 \leqslant x_1, x_2 \leqslant 5 \\ & 0 \leqslant x_i^{+}, x_i^{-}\leqslant 5, \forall i = 1, 2 \end{aligned} maxs.t.2(x1++x1−)+(x2++x2−)x1+2(x2++x2−)⩽8x1=x1+−x1−x2=x2+−x2−(可以删去)−5⩽x1,x2⩽50⩽xi+,xi−⩽5,∀i=1,2
这个是不等价的。可以自行验证。这样的话,会出现 x 1 = 0 , x 1 + = 5 , x 1 − = 5 x_1 = 0, x_1^{+} =5, x_1^{-}=5 x1=0,x1+=5,x1−=5, 也满足 x 1 = x 1 + − x 1 − x_1 = x_1^{+} - x_1^{-} x1=x1+−x1−。
总结
含有绝对值形式的线性化。
- 情况1: 目标函数为 min \min min,例如
min ∣ x ∣ x ⩾ 0 \begin{aligned} \min \,\,\,& |x| \\ &x \geqslant 0 \end{aligned} min∣x∣x⩾0
则可以引入辅助变量 y y y,用下面的形式等价线性化
min y s . t . y ⩾ x y ⩾ − x \begin{aligned} \min \,\,\, y \\ s.t. \,\,\, &y \geqslant x \\ &y \geqslant -x \end{aligned} minys.t.y⩾xy⩾−x
- 情况2: 目标函数为 max \max max
- 例如下面
max 2 ∣ x 1 ∣ + ∣ x 2 ∣ s . t . ∣ x 1 ∣ + 2 ∣ x 2 ∣ ⩽ 8 − 5 ⩽ x 1 , x 2 ⩽ 5 \begin{aligned} \max \quad \,\,\,& 2|x_1| + |x_2| \\ s.t. \quad \,\,\,& |x_1| + 2|x_2| \leqslant 8 \\ & -5 \leqslant x_1, x_2 \leqslant 5 \end{aligned} maxs.t.2∣x1∣+∣x2∣∣x1∣+2∣x2∣⩽8−5⩽x1,x2⩽5
则需要引入两组辅助变量 x i + , x i − , ∀ i = 1 , 2 x_i^{+}, x_i^{-}, \forall i =1,2 xi+,xi−,∀i=1,2。其中
x i + = max { 0 , x i } x i − = max { 0 , − x i } \begin{aligned} &x_i^{+} = \max\{0, x_i\} \\ &x_i^{-} = \max\{0, -x_i\} \end{aligned} xi+=max{0,xi}xi−=max{0,−xi}
因此,就有
x i = x i + − x i − ∣ x i ∣ = x i + + x i − \begin{aligned} &x_i = x_i^{+} - x_i^{-} \\ &|x_i| = x_i^{+} + x_i^{-} \end{aligned} xi=xi+−xi−∣xi∣=xi++xi−
最终变化为
max 2 ( x 1 + + x 1 − ) + ( x 2 + + x 2 − ) s . t . ( x 1 + + x 1 − ) + 2 ( x 2 + + x 2 − ) ⩽ 8 x 1 = x 1 + − x 1 − x 2 = x 2 + − x 2 − x 1 abs = x 1 + + x 1 − x 2 abs = x 2 + + x 2 − x 1 abs ⩾ x 1 x 1 abs ⩾ − x 1 x 2 abs ⩾ x 2 x 2 abs ⩾ − x 2 − 5 ⩽ x 1 , x 2 ⩽ 5 0 ⩽ x i + , x i − , x i abs ⩽ 5 , ∀ i = 1 , 2 \begin{aligned} \max \quad \,\,\,& 2(x_1^{+} + x_1^{-}) + (x_2^{+} + x_2^{-}) \\ s.t. \quad \,\,\,& (x_1^{+} + x_1^{-}) + 2(x_2^{+} + x_2^{-}) \leqslant 8 \\ & x_1 = x_1^{+} - x_1^{-} \\ & x_2 = x_2^{+} - x_2^{-} \\ & x_1^{\text{abs}} = x_1^{+} + x_1^{-} \\ & x_2^{\text{abs}} = x_2^{+} + x_2^{-} \\ & x_1^{\text{abs}} \geqslant x_1 \\ & x_1^{\text{abs}} \geqslant -x_1 \\ & x_2^{\text{abs}} \geqslant x_2 \\ & x_2^{\text{abs}} \geqslant -x_2 \\ & -5 \leqslant x_1, x_2 \leqslant 5 \\ & 0 \leqslant x_i^{+}, x_i^{-}, x_i^{\text{abs}}\leqslant 5, \forall i = 1, 2 \end{aligned} maxs.t.2(x1++x1−)+(x2++x2−)(x1++x1−)+2(x2++x2−)⩽8x1=x1+−x1−x2=x2+−x2−x1abs=x1++x1−x2abs=x2++x2−x1abs⩾x1x1abs⩾−x1x2abs⩾x2x2abs⩾−x2−5⩽x1,x2⩽50⩽xi+,xi−,xiabs⩽5,∀i=1,2
注意:在一些特定 的情况下,表示绝对值的辅助变量
x i abs x_i^{\text{abs}} xiabs也是可以不引入的,也同样可以保证等价。主要看对应的约束和目标函数是否能够在共同的耦合下,使得
x 1 = 0 , x 1 + = 5 , x 1 − = 5 x_1 = 0, x_1^{+} =5, x_1^{-}=5 x1=0,x1+=5,x1−=5,此时
x 1 = x 1 + − x 1 − x_1 =x_1^{+} - x_1^{-} x1=x1+−x1−,但是 ∣ x 1 ∣ ≠ x 1 + + x 1 − |x_1| \ne x_1^{+} + x_1^{-} ∣x1∣=x1++x1−的这种情况出现。
欢迎关注我们的微信公众号 运小筹
公众号往期推文如下