Stochastic dual dynamic integer programming (SDDiP)
Introduction
Multistage stochastic programming
- The resulting formulation, called the extensive form, is
- An alternative to the extensive form (1.1) is to formulate the MSP problem via the following dynamic programming (DP) recursions
MSIP with binary state variables
-
In this section, we
- describe the class of MSIP problem we consider in this work
- propose the key reformulation
-
一些假设:
- 线性(本文结果可以很容易地拓展到非线性情况,但本文聚焦于线性情况)
- 一个重要的要求!!!——
x
n
x_n
xn要是binary的,
y
n
y_n
yn无所谓
之所以有这个要求是因为:
- complete continuous recourse (can always be achieved by adding nonnegative auxiliary con- tinuous variables and penalizing them in the objective function)
- 对本文至关重要的对(1.1)的reformulation——引入
z
n
z_n
zn:
写成DP:
Stochastic nested decomposition (SND) and SDDiP
- In this section, we present SND and its special case SDDiP for solving (2.1):
- SDDiP --> when the stochasticity satisfies stage-wise independence
The SND algorithm
- a sampling step
- a forward step
- a backward step
The goal of the backward step is to update the approximate expected cost-to-go function ψ n i ( ⋅ ) \psi_n^i(\cdot) ψni(⋅) for each sampled node n ∈ ω k i n \in \omega_k^i n∈ωki.
The SDDiP algorithm
- for the setting where the scenario tree satisfies stage-wise independence, i.e., for any two nodes n n n and n ′ n' n′ in S t S_t St the set of children nodes C ( n ) \mathcal{C}(n) C(n) and C ( n ′ ) \mathcal{C}(n') C(n′) are defined by identical data and conditional probabilities.
-
Q
n
(
⋅
)
≡
Q
t
(
⋅
)
\mathcal{Q}_{n}(\cdot) \equiv \mathcal{Q}_{t}(\cdot)
Qn(⋅)≡Qt(⋅) for all
n
∈
S
t
n\in S_t
n∈St
注:随机性出现在约束的左边或者目标函数中较难处理,会出现bilinear项
Sufficient cut conditions
- SND和SDDiP算法是valid的只要cut满足下面三个性质:
Convergence
-
本节证明了SND算法的收敛性
- with probability one, the approximate cost-to-go functions constructed using valid, tight, and finite cuts define an optimal solution to MSIP with binary state variables in a finite number of iterations
-
一个technical假设(大多数确定性MIP求解器都满足,如CPLEX和Gurobi)
-
核心定理:
Cut families
Benders’ cut
不tight
Integer optimality cut
三个性质都满足,但只在指定的解上是紧的,在其他解上非常松 ⇒ \Rightarrow ⇒收敛很慢
Lagrangian cut
- obtained by solving Lagrangian dual of the reformulated nodal problems
- 证明Lagrangian cuts满足三条性质:valid, tight and finite