Two-Stage Robust Optimization (TSRO)

• Two-Stage Robust Optimization (TSRO) is a decision framework that splits actions into pre-uncertainty 'here-and-now' and post-uncertainty 'wait-and-see' stages.
• TSRO generalizes static robust optimization by incorporating adjustable recourse, enabling less conservative and more flexible solutions.
• Its mathematical formulation uses a min-max-min structure and leverages techniques like K-adaptability to manage intractable uncertainty sets.

Two-Stage Robust Optimization (TSRO) is a framework in mathematical optimization wherein decisions are staged: a set of "here-and-now" decisions is made prior to the realization of uncertainty, followed by "wait-and-see" reactions after uncertain parameters have been revealed. This paradigm is central for modeling practical settings in which uncertainty must be addressed proactively, yet certain adjustments are possible once more information is available. TSRO generalizes static robust optimization by allowing for adjustable recourse, thereby accommodating less conservative and more flexible solutions than the static case.

1. Mathematical Formulation and Foundations

A general two-stage robust optimization problem is written as

$$\min_{x\in X}\;\max_{z\in Z}\;\min_{y\in Y} \left\{\,c(z)^\top x + d(z)^\top y:\;T(z)x + W(z)y \le h(z)\,\right\},$$

where $x$ (first stage) is decided before uncertainty is realized, $z$ indexes the uncertainty set (often high-dimensional), and $y$ (second stage) are recourse actions executed after observing $z$ (Julien et al., 2022). This min-max-min structure captures the adversarial nature of robust problems and the sequential opportunity for recourse. In integer and mixed-integer variants,

$$x \in X \subseteq \{0,1\}^n, \quad y \in Y \subseteq \{0,1\}^m,$$

and all components of the data (objective, constraints) may vary with $z$ (Dumouchelle et al., 2023).

Because the set $Z$—and thus the set of all possible "responses"—is typically large or even infinite, (and the recourse function is often intractable), direct solution methods are rarely practical unless specific structure is exploited.

2. Solution Approaches: Algorithms and Approximations

2.1. $K$-Adaptability and Branch-and-Bound

A canonical tractable relaxation is the $K$-adaptability scheme, which pre-specifies $K$ candidate recourse policies $y_1, ..., y_K$. For each realization, the best available $y_k$ is implemented: \