Causal-Bandit Inference for Computer Agriculture Herbicide-Rotation Timing Under Resistance-Feedback Dynamics: An Artificial Intelligence Regret-Bound Evaluation

Weed resistance to herbicides imposes a sequential decision problem on arable farming: each application both controls the current weed population and selects for resistance that degrades the value of future options. the present analysis formalises herbicide-rotation timing as a causal bandit in which the reward of an action depends on the counterfactual resistance state that would have obtained under alternative histories, and in which the resistance state itself evolves as a feedback process. We derive a regret bound for a causal upper-confidence-bound policy that separates the direct effect of an application on yield from its resistance-feedback effect, and we characterise the conditions under which the bound is tighter than that of a resistance-blind bandit. The bound is expressed in terms of the number of rotations, the resistance-feedback coefficient, the variance of yield outcomes and the causal identification gap. A design and power analysis, based on a before-after-control-impact estimator with a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64), indicates that 54 clusters are required to detect a 5-percentage-point effect at 80% power across 2000 replications per design cell. The analysis is theoretical and methodological; no observational data are used or claimed. The contribution is a regret bound that makes the cost of ignoring resistance feedback explicit and a design analysis that quantifies the cluster counts needed to evaluate causal bandit policies in computer agriculture.