Irrigation scheduling under nonstationary rainfall is a sequential decision problem in which the value of information decays with forecast horizon and the cost of regret accumulates across the season. The framework is operationalised as a regret-decomposition evaluation protocol for comparing irrigation policies without requiring a single stationary generative model. Using NASA POWER monthly reanalysis for Kampala, Uganda (2000–2023, 312 observations), the paper reports a mean air temperature of 22.443 °C and a decadal trend of −0.1564 °C, which is used to calibrate the nonstationarity component of the regret bound. The dominant source of regret in short-horizon irrigation scheduling is estimation regret rather than nonstationarity regret, and that this ordering reverses as the forecast horizon lengthens. The contribution is a decomposition that makes the trade-off between information acquisition and policy robustness explicit, together with a reproducible evaluation protocol that separates the regret attributable to learning from the regret attributable to environmental change.
No takes yet. Share an insight, caveat, or question.
Tarr et al. (2026) studied this question.