Abstract:Fluid models provide a tractable approach to approximate multiclass processing networks. This tractability is a due to the fact that optimal control for such models is a solution of a Separated Continuous Linear Programming (SCLP) problem. Recently developed revised SCLP-simplex algorithm allows to exactly solve very large instances of SCLPs in a reasonable time. Furthermore, to deal with the inherent stochasticity in arrival and service rates in processing networks, robust optimization approach is applied to SCLP models. However, a robust counterpart of SCLP problem has two important drawbacks limiting its tractability. First, the robust counterpart of SCLP problem is a huge SCLP problem itself, that can be in several orders of magnitude bigger then the nominal SCLP problem. Second, robust counterpart of SCLP is a degenerate optimization problem, that is not suitable for revised SCLP-simplex algorithm. In this paper we develop theoretical results and a corresponding algorithm that allows to preserve dimensions of nominal SCLP problem and avoid degeneracy issues during solution of its robust counterpart.
What problem does this paper attempt to address?
This paper attempts to solve the robust optimization problem of fluid models in processing networks. Specifically, the authors focus on how to effectively solve the Separated Continuous Linear Programming (SCLP) problem with uncertainty, especially in the face of the inherent randomness in arrival and service rates. The main contribution of the paper lies in the development of a new algorithm and theoretical results to overcome two major challenges encountered by existing methods in solving the uncertain SCLP problem:
1. **Scale problem**: The robust counterpart (RC) of the uncertain SCLP problem is usually several orders of magnitude larger than the nominal SCLP problem, which makes it very difficult to solve.
2. **Degeneracy problem**: The robust counterpart of the uncertain SCLP problem is a degenerate optimization problem and is not suitable for solution using the existing revised SCLP - Simplex algorithm.
To solve these problems, the authors propose a new method, including:
- **Reduction method**: Convert part of the budget uncertainty set into a box - type uncertainty set, thereby eliminating the need for additional variables or constraints.
- **Cutting - plane algorithm**: A cutting - plane algorithm specifically designed for uncertain SCLP, which can be seamlessly integrated with the revised SCLP - Simplex algorithm with minimal modification.
Through these improvements, the authors provide a more efficient method for solving the uncertain SCLP problem, especially when the service rate belongs to the budget uncertainty set. This method not only retains the dimension of the nominal SCLP problem but also avoids the degeneracy problem, thereby improving the solution efficiency and accuracy.
### Summary of Mathematical Formulas
1. **Original Optimization Problem (1)**:
\[
\max_{\eta(t), x(t)} \int_0^T (\gamma^T+(T - t)c^T)(\eta(t)\circ\mu(\xi(t)))dt
\]
Subject to:
\[
\begin{aligned}
&\int_0^t G(\eta(s)\circ\mu(\xi(t)))ds + x(t)=\alpha+at,\\
&H\eta(t)\leq b,\\
&x(t),\eta(t)\geq0,\quad0\leq t\leq T,\quad\xi(t)\in U.
\end{aligned}
\]
2. **Robust Dual Problem (2)**:
\[
\max_{\eta,\beta,\gamma,y}\int_0^T (T - t)c^T\eta(t)-y(t)dt
\]
Subject to:
\[
\begin{aligned}
&\sum_{j = 1}^J\int_0^t G_{k,j}\eta_j(s)ds+\sum_{i = 1}^I\Gamma_i\beta_{k,i}(t)+\sum_{i = 1}^I\sum_{j:s(j)=i}\gamma_{k,i,j}\leq\alpha_k+a_kt,\quad\forall k,t,\\
&\beta_{k,i}(t)+\gamma_{k,i,j}(t)\geq\int_0^t\tilde{G}_{k,j}\eta_j(s)ds,\quad\forall t,k,i,j:s(j)=i,\\
&y(t)\geq\sum_{i = 1}^I\left(\Gamma_i\beta_{0,i}(t)+\sum_{j:s(j)=i}\gamma_{0,i,j}(t)\right),\\
&\beta_{0,i}(t)+\gamma_{0,i,j}(t)\geq\int_0^t
\end{aligned}
\]