A Conservative Stochastic Contextual Bandit Based Framework for Farming Recommender Systems

Shana Moothedath, Xian Yeow Lee, Talukder Jubery, Baskar Ganapathysubramanian, Soumik Sarkar

Keywords: Multi-arm Bandits, Recommender systems, Precision farming

TL;DR: This work presents a contextual bandit framework for farming recommendation system whose objective is to recommend crops/seeds, fertilizers, and irrigation to a farm under a given context in order to maximize the overall net profit.

Abstract: The goal of this work is to develop a recommendation system for smart farming such that based on the details of the farm and the farming conditions, including information on the weather and soil properties, the system presents recommendations on the choice of the crop/seed, the type of fertilizers, and the amount of water for irrigation in order to maximize the overall net profit of the farmer. We refer to this problem as the crop-fertilizer-irrigation (CFI) recommender problem. In this paper, we propose a conservative, stochastic, contextual bandit formulation for solving the CFI problem, where the context captures the farm ID, weather indices and soil properties, the action set is the possible types of crops/seed, the types of fertilizers and irrigation, and the reward is the net profit of the farmer. Our bandit formulation is a conservative bandit setting since we incorporate constraints on the learned policy such that the learned policy need to satisfy certain performance criteria imposed by the farmer while maximizing the reward. Furthermore, our bandit formulation is also stochastic in the sense that the contexts are not ob- servable, rather a distribution of the contexts are known. The stochastic bandit setting captures the uncertainty associated with the measurements of the weather indices and the soil properties. Based on the optimism in the face of uncertainty principle, we propose an algorithm to solve the bandit formulation of the CFI problem. To validate the performance of our approach we used the maize data collected through the G2F initiative. While we present initial model fitting results, the implementation and validation of the proposed algorithm are part of our future work.