contextual bandit policy improves profit in offline evaluation
Open Access DepositedCausal machine learning for fertilizer recommendations
This praxis develops and validates a causal machine learning framework foroptimizing fertilizer recommendations in smallholder maize systems, using historical farm data from Chiapas, Mexico. Smallholder yields in the region remain below potential due to a lack of fertilizer guidelines that take into account heterogeneity in soil conditions, climate, and management. To address this gap, the study formulates fertilizer recommendation as a one-step offline contextual bandit problem, integrating ensemble surrogate crop modeling with conservative policy optimization and robust off-policy evaluation. A multi-year dataset (2012–2018) comprising 4,585 maize field observations wasused to train a stacked ensemble surrogate model – combining XGBoost, LightGBM, and CatBoost learners with a ridge meta-learner – to predict profit responses to nitrogen, phosphorus, and potassium (N–P₂O₅–K₂O) applications under varying conditions. This model achieved an average R² of 0.65 and an RMSE of 3,561 MXN/ha on unseen data, demonstrating strong predictive performance in a highly variable agronomic and economic context. The surrogate serves as the foundation for a conservative contextual bandit policy constrained to historically-supported fertilizer regimes, ensuring that its fertilizer recommendations maintain agronomic and empirical plausibility. Policy performance was evaluated offline using doubly robust (DR) and self-normalized doubly robust (SNDR) estimators. Across the evaluation period, the learned policy achieved statistically significant profit improvements over the historical baseline while satisfying pre-specified support criteria and exhibiting well-behaved importance weights – confirming reliability of the offline estimates. These results show that conservative, data-driven policies can extract economically meaningful insights from observational agronomic data without new field experiments. The validated policy was operationalized through a prototype bilingual decision-support web application that translates the model into interpretable, site-specific N–P₂O₅–K₂O recommendations with predicted profit outcomes. This praxis thus bridges the gap between machine learning research and practical agricultural decision support, demonstrating how conservative causal and reinforcement-learning principles can enhance smallholder profitability and input efficiency.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.