运筹学 $5约束极值

§5 约束极值

C1 最优性条件

1)非起作用约束:当前可行解不在约束边界的约束

g ( X 0 ) > t g(X_0)\gt t g(X0)>t满足 g ( X 0 ) ≥ t g(X_0)\ge t g(X0)t但不在边界上

2)可行方向 ∃ λ : X 0 + λ D ∈ R    ⟺    \exist\lambda:X_0+\lambda D\in R\iff λ:X0+λDR对于所有起作用约束 ∇ g ( X 0 ) T D ≥ 0 \nabla g(X_0)^TD\ge 0 g(X0)TD0

3)下降方向:若 ∃ δ ∀ λ ∈ [ 0 , δ ] f ( X 0 + λ D ) < f ( X 0 )    ⟺    ∇ f ( X 0 ) T D < 0 \exist \delta \forall \lambda\in[0,\delta]f(X_0+\lambda D)\lt f(X_0)\iff \nabla f(X_0)^TD\lt0 δλ[0,δ]f(X0+λD)<f(X0)f(X0)TD<0

4)可行方向法:解 min ⁡ η { ∇ f ( X k ) T D < η ∇ g i ( X k ) T D > η − 1 ⪯ D ⪯ 1 \min \eta\begin{cases}\nabla f(X_k)^TD\lt \eta\\\nabla g_i(X_k)^TD\gt \eta\\ -1\preceq D\preceq 1\end{cases} minηf(Xk)TD<ηgi(Xk)TD>η1D1

  • η < 0 \eta\lt 0 η<0,则对应最优可行下降方向
  • ∣ η ∣ < ϵ |\eta|\lt \epsilon η<ϵ,可停止计算

5)K-T条件:约束极值点满足 { ∇ f ( X ∗ ) − ∑ i = 1 m γ i ∇ g i ( X ∗ ) = 0 γ i g i ( X ∗ ) = 0 γ i ∗ ≥ 0 \begin{cases}\nabla f(X^*)-\sum\limits_{i=1}^m\gamma_i\nabla g_i(X^*)=0\\\gamma_i g_i(X^*)=0\\\gamma_i^*\ge0\end{cases} f(X)i=1mγigi(X)=0γigi(X)=0γi0

  • 对于凸规划,是充要条件;对于线性规划,是必要条件
  • γ i \gamma_i γi称为广义拉格朗日乘子

C2 二次规划

1)二次规划:目标函数为二次形式,且约束条件线性
min ⁡ f ( X ) = ∑ j = 1 n c j x j + 1 2 ∑ j = 1 n ∑ k = 1 n c j k x j x k , c j k = c k j { ∑ j = 1 n a i j x j + b i ≥ 0 x j ≥ 0 \min f(X) =\sum\limits_{j=1}^n c_jx_j + \frac{1}{2}\sum\limits_{j=1}^n\sum\limits_{k=1}^n c_{jk}x_jx_k,c_{jk} = c_{kj}\\ \begin{cases} \sum\limits_{j=1}^n a_{ij} x_j +b_i\ge 0\\ x_j\ge 0 \end{cases} minf(X)=j=1ncjxj+21j=1nk=1ncjkxjxk,cjk=ckjj=1naijxj+bi0xj0
2)K-T解法:问题等价于线性规划问题
min ⁡ ϕ ( Z ) = ∑ j = 1 n z j { ∑ i = 1 m a i j γ n + i + γ i − ∑ k = 1 n c j k x k + s g n ( c j ) z j = c j ∑ j = 1 n a i j x j − x n + i + b i = 0 x j ≥ 0 , y j ≥ 0 , z j ≥ 0 \min \phi(Z) = \sum\limits_{j=1}^n z_j\\ \begin{cases} \sum\limits_{i=1}^m a_{ij} \gamma_{n+i} +\gamma_i -\sum\limits_{k=1}^n c_{jk}x_k+sgn(c_j)z_j=c_j\\ \sum\limits_{j=1}^n a_{ij}x_j - x_{n+i}+b_i=0\\ x_j\ge 0, y_j \ge 0,z_j \ge 0 \end{cases} minϕ(Z)=j=1nzji=1maijγn+i+γik=1ncjkxk+sgn(cj)zj=cjj=1naijxjxn+i+bi=0xj0,yj0,zj0

  • 初始可行解

{ z j = s g n ( c j ) c j x n + i = b i x j = y j = 0 \begin{cases} z_j = sgn(c_j)c_j\\ x_{n+i}=b_i\\ x_j=y_j=0 \end{cases} zj=sgn(cj)cjxn+i=bixj=yj=0

C3 制约函数法

1)外点法

  • ϕ ( g i ( X ) ) = { 0 , t ≥ 0 t 2 , t < 0 \phi(g_i(X))=\begin{cases} 0,t\ge 0\\t^2,t\lt 0\end{cases} ϕ(gi(X))={0,t0t2,t<0惩罚函数 P ( X , M ) = f ( X ) + M ∑ i = 1 l ϕ ( g i ( x ) ) P(X,M)=f(X)+M\sum\limits_{i=1}^l \phi(g_i(x)) P(X,M)=f(X)+Mi=1lϕ(gi(x))
  • 求解无约束极值问题 min ⁡ P ( X , M k ) \min P(X,M_k) minP(X,Mk)
  • 若某个 g i ( X ) < − ϵ g_i(X)\lt -\epsilon gi(X)<ϵ,则增大M,求解 min ⁡ P ( X , M k + 1 ) \min P(X,M_{k+1}) minP(X,Mk+1)。多次迭代直至得到可行解

2)内点法

  • g ( X , r ) = r k ∑ j = 1 l 1 g j ( X ) g(X,r)=r_k\sum\limits_{j=1}^l\frac{1}{g_j(X)} g(X,r)=rkj=1lgj(X)1 g ( X , r ) = − r k ∑ j = 1 l log ⁡ ( g j ( X ) ) g(X,r)=-r_k\sum\limits_{j=1}^l \log(g_j(X)) g(X,r)=rkj=1llog(gj(X))障碍函数 P ˉ ( X , r ) = f ( X ) + g ( X , r ) \bar P(X,r)=f(X)+g(X,r) Pˉ(X,r)=f(X)+g(X,r)

  • 取一可行解 X 0 X_0 X0为初始点,求解无约束极值问题 min ⁡ P ˉ ( X , r k ) \min \bar P(X,r_k) minPˉ(X,rk)

  • g ( X , r k ) ≤ ϵ g(X,r_k)\le \epsilon g(X,rk)ϵ,则终止。否则减小r,计算 min ⁡ P ˉ ( X , r k + 1 ) \min \bar P(X,r_{k+1}) minPˉ(X,rk+1)

©️2020 CSDN 皮肤主题: 大白 设计师:CSDN官方博客 返回首页