Off-Policy Evaluation for Artificial Intelligence Agriculture Irrigation Scheduling Under Unobserved Evapotranspiration Feedback: A Computer Doubly-Robust Bound

Off-policy evaluation (OPE) of irrigation scheduling policies is complicated by unobserved evapotranspiration (ET) feedback: the same soil-moisture state that a policy acts upon is also shaped by the policy's own past actions, so that the logging distribution and the target distribution differ in ways that standard importance-sampling estimators cannot correct. The bound is derived under a sequential ignorability assumption conditional on a proxy for ET, and it is shown to be sharp when the proxy is informative and to degrade gracefully when it is not. The derivation yields a finite-sample concentration inequality that accounts for the variance of the importance weights and the bias of the outcome model, and it produces a computable upper bound on the mean squared error of the OPE estimate. The bound is illustrated with a reanalysis of a public reanalysis dataset for Nairobi, Kenya (2000–2023, 312 monthly observations), in which the mean temperature is 20.146 °C and the warming trend is 0.2477 °C per decade; these values are used to parameterise the ET proxy and to show how the bound responds to a realistic warming signal. The contribution is a computer-science result: a bound that is tight enough to be useful in practice, computable from logged data alone, and explicit about the conditions under which OPE of irrigation policies can be trusted.