Robust Optimisation

Robust Optimisation

Robust optimisation chooses a feasible decision that performs acceptably across a specified set of uncertain parameter values. In its classic form, the chosen decision must satisfy required constraints for every value in that uncertainty set, and the objective is evaluated at its worst permitted value. This makes protection explicit, but it is only as credible as the uncertainty set, cost function, constraints and information available when the decision is made.

Decision, uncertainty and timing

Let $x$ be a decision selected now and $u$ an uncertain vector known only to belong to set $\mathcal{U}$. For a minimisation problem, a common static robust formulation is

$$\min_{x\in X}\ \max_{u\in\mathcal{U}} C(x,u)\quad\text{subject to}\quad g_j(x,u)\leq0\ \text{for all }u\in\mathcal{U},\ j=1,\ldots,m.$$

Here $X$ contains design or resource constraints known in advance, $C$ is loss or cost, and $g_j$ are requirements that must hold under every included scenario. For a health-gain objective, the equivalent perspective is to maximise the minimum gain across the set. The max or min within the objective is not a forecast of the most likely outcome; it identifies the adversarial case permitted by the modelled set.

Static decisions are fixed before uncertainty resolves. In a two-stage or adjustable robust model, some actions can adapt after part of $u$ is observed—for example, reserve capacity now and allocate staff later. Specify the information revealed at each stage and the permitted recourse; allowing a decision to change after seeing demand while calling it fixed would overstate attainable performance. A robust constraint is a guarantee within the chosen set and model, not a guarantee against every future event.

Constructing an uncertainty set

An interval set may place each uncertain demand between defensible lower and upper bounds; a finite scenario set lists specific combinations. A box that allows all inputs simultaneously to reach their adverse extremes may be unrealistically conservative if the extremes cannot occur together. Budgeted uncertainty, correlated ellipsoids and data-informed sets can restrict joint deviations, but they introduce parameters and assumptions that need explanation.

Specify source data, horizon, units, dependence, excluded scenarios and the level of conservatism. A confidence region or prediction region can inform a set, yet a deterministic robust guarantee over that set is not automatically a probability guarantee unless the set's coverage and the probability model are separately justified. Changes in epidemiology, service delivery or prices can make a historical set incomplete. Test the preferred decision as the set expands or contracts.

Worked health service capacity example

A clinic must choose daily capacity $x\in\{80,100,120\}$ before learning tomorrow's demand $d\in\mathcal{U}=\{80,100,120\}$. Every slot costs 10 currency units, and every patient above capacity incurs a modelled shortage loss of 30 units. Assume unused capacity has no other cost beyond the slot cost and all three choices meet an existing 120-slot ceiling. The invented objective is

$$C(x,d)=10x+30\max(d-x,0).$$

At the nominal demand of 100, capacity 100 costs 1,000 and is the least costly listed choice. Across the entire specified uncertainty set, capacity 120 has the smallest worst cost, 1,200, so it is the robust minimax choice. It costs 200 more than the nominal optimum when demand is 100, and improves worst-set cost by 400 compared with capacity 100. These are two different comparisons; neither is an expected cost without probabilities for demand.

The shortage loss is a modelling assumption that may not reflect the true health consequences of delayed care. If capacity 120 were unaffordable, infeasible, or crowded out higher-value services, it could not be chosen under the revised constraints. If demand exceeded 120, the calculated within-set protection would no longer apply. The analysis should therefore vary the uncertainty set, unit costs and shortage valuation, and disclose whether demand is measured daily, averaged or correlated across days.

Robust counterpart and the price of protection

For a finite scenario set, the outer minimax can be written with an auxiliary upper bound $z$: minimise $z$ subject to $z\geq C(x,u)$ for every $u\in\mathcal{U}$, together with robust feasibility constraints. Continuous sets require mathematical reformulation or algorithms appropriate to their shape and the model's convexity. A computational solution should report feasibility and optimality tolerances, and the adverse scenarios that make constraints bind.

The “price of robustness” describes how protection changes the objective under a stated reference comparison. In the example, choosing 120 instead of the nominal-optimal 100 sacrifices 200 units at demand 100 to reduce worst-set cost by 400. A different uncertainty budget can yield a different balance between nominal performance and protection. Never present the robust choice as best in every scenario; at demand 80, capacity 80 costs 800 versus 1,200 for capacity 120.

Robust optimisation is distinct from one-way sensitivity analysis, which probes a fixed decision or model result by changing inputs. It is also distinct from stochastic programming, which optimises an expected outcome under an explicit probability distribution, and from chance-constrained models, which permit a specified probability of constraint violation under such a distribution. Distributionally robust optimisation takes a worst case over a set of probability distributions, a further modelling choice that should be named rather than folded into an ordinary parameter uncertainty set.

Health economic interpretation

Applications include emergency capacity, medication stock, workforce planning and allocations when need, cost or treatment effect is uncertain. The objective may combine health loss and resource cost, or keep clinical access requirements as hard constraints. Put units on each term and document any conversion of health consequences into money or another common scale. A worst-case cost rule can conflict with equity if disadvantaged groups are hidden within aggregate demand; add distributional objectives or constraints when the decision requires them.

Robust cost-effectiveness decisions can compare net benefit under parameter ranges, but the set of alternatives, budget constraints and willingness-to-pay threshold remain essential. A strategy with the highest worst-case net benefit need not have the highest expected net benefit, and a robustly feasible plan need not have a positive expected net benefit. Decision makers should see scenario outcomes, trade-offs and opportunity costs rather than a single recommended solution without its assumptions.

Validation and reporting

Check the input bounds and joint plausibility, decision timing, feasibility in each stress scenario, objective calculations, and sensitivity to wider sets. Compare nominal, robust and plausible alternative decisions under common evaluation scenarios. Out-of-sample stress tests or prospective monitoring can reveal where the assumed set fails, though passing historical tests does not establish future protection. Report which constraint or scenario is binding and whether a solution exists if the protection level increases.

Bertsimas and Sim's The Price of Robustness develops budgeted uncertainty and explicitly frames protection-versus-performance trade-offs. The Robust Optimization monograph by Ben-Tal, El Ghaoui and Nemirovski presents robust counterparts and uncertainty-set design. A healthcare resource-allocation application illustrates why uncertain demand and capacity matter in practice. The numerical clinic example here is original and illustrative, not a calibrated operational recommendation.