Risk-Aware Bandits for Best Crop Management
Keywords: Bandits; Risk-awareness; Agriculture; on-farm trials
TL;DR: We propose a risk-aware bandit algorithm to help a community of farmer improve their crop management practices
Abstract: Improving fertilizer practices through on-farm trials is challenging, especially in rain-fed farming due to weather uncertainty. However, it is crucial to test various practices to determine their performance, even if some may yield inferior results during the experiment. Our case study focuses on maize production in southern Mali, and we use the Decision Support System for Agrotechnology Transfer (DSSAT) crop model to simulate maize responses to nitrogen fertilization. We compare fertilizer practices using the Conditional Value-at-Risk (CVaR) of the Yield Excess (YE), a novel agronomic metric that considers both grain yield and nitrogen use efficiency. An "intuitive strategy" for practitioners, called *Explore-Then-Commit* (ETC) in the bandit literature, involves multi-year, multi-location field trials, where each practice is tested equally over several years. Inspired by
a recent contribution,
we propose the *Bounded-CVaR TS-Batch* (BCB) bandit algorithm, improving over ETC both theoretically and in crop model simulations. This study opens new horizons for risk-aware identification of best crop management practices' in real conditions.
Submission Number: 105
Loading