A mathematical preprint says the value of human expertise in an uncertain decision problem can be pinned down as the gap between two optimization benchmarks. Under the paper’s stated Assumptions 1 and 2, its main result equates that value with the difference between the optimal values of its min-max and max-min formulations.
At the heart of the work is a question about information that may not appear in a dataset: can a decision maker’s belief that the nominal optimum is unlikely to be large improve a robust decision when parameters are unknown? The study puts that belief into the optimization as a constraint on which parameter scenarios remain possible.
A belief becomes a constraint
The mechanism is described through nominal curves. For a chosen threshold, it excludes parameters that would make the nominal optimum exceed that threshold, then evaluates policies with those nominal curves rather than only their unrestricted worst-case performance. In effect, the model links a belief about the nominal solution to a scenario-specific worst-case guarantee.
The theory gives both a ceiling and, under stronger conditions, an exact measure. Under Assumption 1, the modeled value of expertise is bounded above by the gap between the min-max and max-min formulations. Under Assumptions 1 and 2, that upper bound becomes an equality.
The assortment test
To show how the idea behaves in an application-shaped setting, the paper uses one stylized assortment problem. Its uncertainty set consists of random utility maximization models, which represent uncertain choices among products, consistent with transactional sales data from past assortments. The example assigns per-unit revenues of $2, $10, $31 and $40 to products 1 through 4.
In that case, the strict test for a guaranteed improvement had no solution: the stylized identification problem did not produce an assortment meeting the paper’s strict-improvement criterion. The best past assortment supplied an expected-revenue benchmark of $21.
The policy choice nevertheless shifted with the assumed nominal-value threshold. At about $33.778, S2 is optimal for threshold values at or above that level, while {0, 2, 3, 4} is optimal at or below it when the reduced uncertainty set is nonempty.
When the near-optimality belief is correct, the new assortment’s worst-case expected revenue can be 13.7% higher. If the belief is wrong, the reported worst-case decrease is at most 2.38%. These are conditional guarantees from a stylized numerical example, not observed sales outcomes.
A wider mathematical sweep
The paper then turns to shortest-path problems and tests 20 randomly generated instances. Each instance solves the robust and non-robust problems once, while the reduced problem is solved at 101 equally spaced threshold values using mixed-integer linear programming.
A strictly positive modeled value of human expertise appeared in 19 of the 20 instances. Among those 19, the average ratio between the optimal values was 1.389. Six of the 20 instances also satisfied at least one sufficient structural condition associated with a strictly lower reduced robust value for every allowed threshold in H.
For computation, the paper proposes an exact mixed-integer reformulation for one class of problems and a more general cutting-plane method. These methods are intended to make the reduced policies calculable across different problem types.
What the examples do not show
The results come with a narrow scope. The assortment findings belong to a stylized case, and the shortest-path findings come from randomly generated instances; the paper does not present them as evidence from actual stores, transport networks or human decisions. The reported ratio and percentage changes are therefore model-based guarantees, not population estimates.
Nor does the exact equality stand independently of the assumptions: the paper’s minimax-gap characterization is conditional on Assumptions 1 and 2. The central evidence is thus a mathematical statement about how a belief can change a robust-optimization calculation, not a finding that people have been shown to make better decisions.
The authors’ broader interpretation is that nominal curves can translate informal human or domain knowledge into scenario-specific worst-case guarantees while retaining protection if the belief is wrong. The supplied document is an arXiv v1 preprint dated 26 Aug 2026.
Paper data and sources
Original title: The Value of Human Expertise
Authors: Bradley Sturt
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text