Offline Reinforcement-Learning Inference for Computer Agriculture Irrigation Policies Under Bandwidth-Constrained Edge Deployment: An Artificial Intelligence Regret-Bound Evaluation

Offline reinforcement learning (RL) offers a route to irrigation scheduling that learns from logged field data without online interaction, yet edge deployment under bandwidth constraints forces a choice between transmitting rich state information and executing policies locally on compressed representations. the present analysis characterises the regret incurred by offline RL inference when the policy is trained on a full state representation but deployed on a bandwidth-limited edge node that observes only a quantised subset of state variables. The analysis is theoretical and methodological: no field data are used or claimed. A regret decomposition is derived that separates the irreducible suboptimality of the offline dataset from a representation-mismatch term that scales with the bandwidth budget and the sensitivity of the optimal Q-function to the discarded state coordinates. The bound is evaluated through a replication-based design in which the operating characteristics of the before-after-control-impact estimator are characterised under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64), with power estimated across 2000 design cells. The analysis identifies the bandwidth threshold below which representation mismatch dominates the regret budget, and it shows that the intraclass correlation of 0.64 implies that 52 clusters are required to detect a 5-percentage-point effect at 80% power. The contribution is a deployable regret bound that links communication budget to policy suboptimality, together with a design-analysis protocol that lets irrigation-advice programmes size their edge deployments before committing to a bandwidth contract.