Robust and Distributionally Robust Optimization for WFM

Robust and Distributionally Robust Optimization for WFM addresses a fundamental tension in workforce management: every staffing plan is built on a forecast, and every forecast is wrong. Classical optimization finds the best solution assuming the forecast is exactly right. Robust optimization finds the best solution that still works when the forecast is wrong. This distinction separates fragile staffing plans from resilient ones.

Overview

Uncertainty sets: box, ellipsoidal, and budget around forecast

Workforce management operates under pervasive uncertainty. Contact volume forecasts carry error. Average handle times fluctuate. Attrition rates shift unexpectedly. Absenteeism spikes during flu season. Traditional optimization approaches treat these parameters as known constants, producing solutions that are optimal in a world that doesn't exist.

Three paradigms address this gap:

1. Deterministic optimization — assumes all parameters are known exactly. Fast, tractable, and wrong.
2. Stochastic programming — assumes uncertainty follows a known probability distribution. Better, but requires distributional assumptions that may themselves be wrong.
3. Robust optimization — assumes only that uncertain parameters lie within a set of possible values. Optimizes against the worst case within that set. No distributional assumption required.

Distributionally robust optimization (DRO) occupies the middle ground: it assumes the true distribution belongs to a family (an ambiguity set) and optimizes against the worst-case distribution within that family.

The practical implication for WFM: robust methods produce staffing plans with built-in buffers that are mathematically principled rather than ad hoc. Instead of adding 10% "just in case," robust optimization determines precisely how much buffer is needed given the uncertainty you face.

Mathematical Foundation

Classical Robust Optimization

In standard optimization, we solve:

$\underset{x\in X}{\min }{c}^{T}x\text{subject to}Ax\le b$

where $A$ and $b$ are known with certainty. In robust optimization, the constraint data is uncertain. We define an uncertainty set ${𝒰}$ and require feasibility for every realization:

$\underset{x\in X}{\min }{c}^{T}x\text{subject to}{a}_i^{T}x\le b_i\mathrm{\forall }(a_i,b_i)\in {{𝒰}}_i$

The solution must be feasible regardless of which scenario within ${𝒰}$ materializes. This is the robust counterpart of the original problem.

Uncertainty Set Geometries

The shape of the uncertainty set determines both the conservatism and tractability of the robust counterpart.

Box uncertainty (interval): Each parameter varies independently within a range.

${{𝒰}}_{box}=\{\hat{a}+\delta :|{\delta }_j|\le {\hat{\delta }}_j\mathrm{\forall }j\}$

In WFM terms: "volume could be anywhere from 450 to 550 calls per interval, and AHT anywhere from 280 to 320 seconds." Box uncertainty is the simplest but most conservative — it assumes all parameters can simultaneously hit their worst-case values, which rarely happens.

Ellipsoidal uncertainty: Parameters are correlated, and the uncertainty set forms an ellipsoid.

${{𝒰}}_{ellip}=\{\hat{a}+P\delta :\Vert \delta {\Vert }_2\le \mathrm{\Omega }\}$

where $P$ captures the correlation structure and $\mathrm{\Omega }$ controls the size. This accounts for the fact that if Monday volume is high, Tuesday volume is also likely high (positive correlation).

Budget of uncertainty (Bertsimas & Sim, 2004): The key innovation. Rather than requiring robustness against all parameters hitting worst case simultaneously, a budget parameter $\mathrm{\Gamma }$ controls how many parameters can deviate:

${{𝒰}}_{\mathrm{\Gamma }}=\{\hat{a}+\delta :|{\delta }_j|\le {\hat{\delta }}_j,\sum _j\frac{|{\delta }_j|}{{\hat{\delta }}_j}\le \mathrm{\Gamma }\}$

When $\mathrm{\Gamma }=0$, we get the deterministic solution (no uncertainty). When $\mathrm{\Gamma }=n$ (number of parameters), we get full box uncertainty. The planner tunes $\mathrm{\Gamma }$ to control the trade-off between robustness and cost. This approach is powerful because:

• The robust counterpart remains a linear program (LP) when the original problem is LP
• $\mathrm{\Gamma }$ provides an intuitive knob — "how many things can go wrong simultaneously?"
• Probabilistic guarantees exist: with $\mathrm{\Gamma }=\sqrt{n}$, the probability of constraint violation is bounded

Distributionally Robust Optimization

DRO optimizes against the worst-case distribution within an ambiguity set ${𝒫}$:

$\underset{x\in X}{\min }\underset{P\in {𝒫}}{\sup }{\mathbb{𝔼}}_P[f(x,\xi )]$

Common ambiguity sets include:

• Moment-based: All distributions matching known mean and covariance — ${𝒫}=\{P:{\mathbb{𝔼}}_P[\xi ]=\mu ,{\text{Cov}}_P(\xi )=\mathrm{\Sigma }\}$
• Wasserstein ball: All distributions within a Wasserstein distance $ϵ$ of an empirical distribution — ${𝒫}=\{P:W(P,{\hat{P}}_n)\le ϵ\}$

The Wasserstein approach is particularly useful in WFM because it leverages historical data (the empirical distribution ${\hat{P}}_n$) while hedging against the possibility that history doesn't perfectly predict the future.

WFM Applications

Robust Staffing Plans

The most direct application: generating staffing requirements that are resilient to forecast error.

Standard approach: Forecast 500 calls with AHT 300s, compute staff using Erlang C, add a fixed shrinkage buffer.

Robust approach: Define uncertainty sets around volume (±15%) and AHT (±10%). Use budget-of-uncertainty with $\mathrm{\Gamma }$ calibrated so that no more than 30% of intervals simultaneously exceed forecast. Solve the robust counterpart.

The robust solution produces interval-specific staffing levels that are higher during periods where the consequence of under-staffing is severe (peak hours, contractual SLA windows) and closer to the deterministic optimum during off-peak periods. The buffer is intelligent, not uniform.

Schedule Optimization Under Demand Uncertainty

When building weekly schedules, the optimizer assigns shifts to cover required staffing levels. Robust scheduling replaces point-estimate requirements with uncertain intervals:

$\min \sum _{s\in S}{c}_s{x}_s\text{s.t.}\sum _{s\in S}{a}_{is}{x}_s\ge r_i+{\delta }_i\mathrm{\forall }i,\mathrm{\forall }\delta \in {{𝒰}}_{\mathrm{\Gamma }}$

where $r_i$ is the nominal requirement for interval $i$, ${\delta }_i$ is the demand deviation, $a_{is}$ indicates whether shift $s$ covers interval $i$, and ${{𝒰}}_{\mathrm{\Gamma }}$ is a budget uncertainty set.

Capacity Planning Robust to AHT Variation

Long-term capacity planning is especially vulnerable to parameter uncertainty because decisions (hiring, training) have long lead times and high reversal costs. DRO is natural here: build a Wasserstein ambiguity set from historical monthly volume distributions, then solve the hiring plan that minimizes expected cost under the worst-case distribution within that ball.

Robust Real-Time Optimization

Even intraday adjustments benefit from robustness. When deciding whether to offer VTO at 2 PM, the optimizer faces uncertainty about 3-5 PM volume. A robust formulation keeps enough agents to handle the worst case within the uncertainty budget, preventing the costly error of releasing too many agents before a late-afternoon spike.

Worked Example

Problem: A contact center plans weekly staffing across 96 half-hour intervals per day (7 days = 672 intervals). Forecast volume has a mean absolute percentage error (MAPE) of 12%. The question: how much additional staff does robust optimization recommend compared to the deterministic solution?

Setup:

• Nominal requirement per interval: $r_i$ (from Erlang C calculation)
• Maximum deviation: ${\hat{\delta }}_i=0.15\cdot r_i$ (slightly larger than MAPE to capture tail risk)
• Budget parameter: $\mathrm{\Gamma }=\lceil \sqrt{672}\rceil =26$ (at most 26 of 672 intervals can simultaneously deviate to worst case)

Robust counterpart (Bertsimas & Sim formulation):

For each interval $i$, the robust constraint becomes:

$\sum _s{a}_{is}{x}_s\ge r_i+z_i{\hat{\delta }}_i+p_i$

where auxiliary variables $z_i,p_i$ enforce the budget constraint. The additional decision variables add $O(n)$ variables and constraints — the problem remains an LP.

Results:

• Deterministic solution: 142 FTE, cost index 100
• Robust solution ($\mathrm{\Gamma }=26$): 149 FTE, cost index 104.9
• Full worst-case (box): 163 FTE, cost index 114.8

The robust solution adds approximately 5% cost to protect against realistic uncertainty — far less than the 15% buffer a box uncertainty set would require, and far more principled than an arbitrary "add 10%" rule. Moreover, the additional staff is concentrated in intervals where under-staffing risk is highest (Monday mornings, lunch peaks), not spread uniformly.

Maturity Model Position

Level	Description
Level 1 (Manual)	No uncertainty consideration; staffing based on point forecasts
Level 2 (Developing)	Flat percentage buffers added to forecasts (e.g., "add 10%")
Level 3 (Defined)	Scenario analysis — run optimizer under optimistic, expected, and pessimistic forecasts
Level 4 (Quantitative)	Stochastic programming with scenario trees; formal uncertainty quantification
Level 5 (Optimizing)	Distributionally robust optimization with data-driven ambiguity sets; adaptive robust optimization that adjusts as uncertainty resolves

Most contact centers operate at Level 2. Operations with sophisticated WFM platforms may reach Level 3 through scenario planning. Level 4-5 represents the frontier, primarily seen in large BPO operations and WFM software R&D teams.

Connection to Other Approaches

Robust optimization complements rather than replaces other uncertainty-handling methods:

• Sensitivity analysis tells you which parameters matter most — informing which parameters to include in the uncertainty set
• Monte Carlo simulation can validate robust solutions by sampling from the uncertainty set and verifying feasibility
• Scenario planning is a discrete version of robust optimization (finite uncertainty set)
• Stochastic programming is preferred when distributional assumptions are reliable and the decision-maker is risk-neutral

See Also

• Operations Research in Workforce Management
• Sensitivity Analysis and Duality
• Convex Optimization in Workforce Planning
• Schedule Optimization
• Capacity Planning Methods
• Probabilistic Forecasting
• Multi-Objective Optimization

References

• Ben-Tal, A., El Ghaoui, L., & Nemirovski, A. (2009). Robust Optimization. Princeton University Press.
• Bertsimas, D. & Sim, M. (2004). "The Price of Robustness." Operations Research, 52(1), 35–53.
• Bertsimas, D., Gupta, V., & Kallus, N. (2018). "Data-driven robust optimization." Mathematical Programming, 167(2), 235–292.
• Bertsimas, D. & Tsitsiklis, J.N. (1997). Introduction to Linear Optimization. Athena Scientific.
• Rahimian, H. & Mehrotra, S. (2019). "Distributionally robust optimization: A review." arXiv:1908.05659.