Stochastic vs. Robust: When to Use Which

Decision criteria for choosing between stochastic and robust optimization, based on (1) what is known about the uncertain data — a full distribution vs. only an uncertainty set — and (2) what the decision-maker cares about — average performance vs. worst-case guarantee. Includes the price of robustness and the role of hard constraints in forcing the robust framework.

Step 1 of 157%

Tutorial

The Two Paradigms

When an optimization problem has uncertain data ξ\xi, two paradigms are available.

Stochastic optimization treats ξ\xi as a random variable with a known distribution PP, and minimizes expected cost:

min⁡x∈X  Eξ∼P ⁣[f(x,ξ)].\min_{x \in X} \; \mathbb{E}_{\xi \sim P}\!\left[f(x,\xi)\right].

Robust optimization treats ξ\xi as an unknown element of a deterministic uncertainty set U\mathcal{U}, and minimizes worst-case cost:

min⁡x∈X  max⁡ξ∈U f(x,ξ).\min_{x \in X} \; \max_{\xi \in \mathcal{U}}\, f(x,\xi).

The two paradigms answer different questions:

  • Stochastic asks what is best on average?
  • Robust asks what is best in the worst case?

Choosing between them comes down to two questions:

  1. What do we know about ξ\xi? — A full distribution, or only a set of possibilities?
  2. What do we care about? — Average cost, or a guarantee against the worst case?

Both questions can point to the same framework or to different ones. When they conflict, the higher-stakes question wins.

↑ ↓ navigate · Enter open · Esc close · ⌘K/Ctrl K toggle