Robust decision-making

Lecture 04

September 29, 2025

Predict-Then-Act vs. Agree-On-Decisions

The DMDU Argument Against Predict-Then-Act

wicked problems are not well-bounded, are framed differently by various stakeholders, and are not well-understood until after formulation of a solution. Using predictions to adjudicate such problems skews attention toward the proverbial lamp post, not the true location of the keys to a policy solution.

— Lempert, 2019. Robust Decision Making (RDM).

Finding robust solutions that “perform well” over a “wide range” of futures

Rob Lempert, PhD; & Michelle Miro, PhD. Online Training for Water Utilities—WUCA / Chapter 3: Plan

Measuring robustness

Regret: The “cost” of choosing “incorrectly”

Adaptating notation in Lempert and Collins (2007), the regret of action a in state x is:

\(R_{a}(x) = Max_{a'}[PV(U_{a'}(x))] - PV(U_{a}(x))\),

the difference between the present value expected utility of the optimal action in a considered state of the world and the considered action’s present value expected utility in that state of the world.

Note: There are many ways to go from an action’s regret metric in one state (i.e., \(T_{1}\)) to evaluating its overall robustness (i.e., \(T_{2}\) and \(T_{3}\)).

Satisficing

Following McPhail et al. (2019), this approach “seeks to meet a minimum performance threshold.”

Looking across states (i.e., \(T_{2}\) and \(T_{3}\)), the domain criterion is a common approach (e.g., see Herman et al., (2015)), which analysts often implement as the fraction of states of the world in which the performance threshold is met.