Benchmarking Evaluation of Offline Reinforcement Learning for Computer Agriculture Irrigation Control Under Logged-Policy Confounding: An Artificial Intelligence Pessimism Bound for Water-Use Safety

Offline reinforcement learning (RL) promises irrigation controllers that can be trained from logged field data without further water expenditure, but the logged policies that generated those data confound every value estimate the controller relies on. The bound is derived from a concentration inequality on importance-weighted returns, corrected for the behaviour-policy overlap that determines how much of the target policy's state-action distribution the logs actually cover. The resulting penalty is a function of the effective sample size of the logged dataset and the estimated confoundedness of the behaviour policy, and it yields a certified lower confidence bound on the water-use safety of any candidate irrigation policy. The bound is validated against a real World Bank panel of 24 African countries observed from 2000 to 2023 (N = 560), in which GDP per capita (constant 2015 US$) was regressed on access to electricity and urban population share; the fitted model (R-squared = 0.591, F = 184.764) supplies the macro-contextual covariate structure against which the irrigation benchmark is calibrated. Offline RL benchmarking for irrigation must report the pessimism penalty alongside the policy return, because a controller that is safe only under the logging policy is not safe at all.