Summary of DC4SCM project
World’s population is expected to reach almost 10 billions by 2050 (FAO). Consumpution of meat, fruits, vegetables, … will increase because of the growth of the population, but also because of a global increase of income of the population. Agriculture as mainly practiced in western countries is unsustainable, and its ability to increase the yields is slowing down. It is then important to study and develop new ways for agriculture that are sustainable and able to meet the population growth and its needs, respect and preserve nature, cope with the big challenge of the global warming. This project, DC4SCM, is obviously extremly modest with regards to these huge planetary challenge, but as modest as it is, DC4SCM aims at going towards new directions of thinking and making agriculture.
Crop-management can be defined as the “logical and ordered combination of techniques applied to a plot aiming at a production” (Sebillotte, 1974). This definition emphasizes the sequential nature of agricultural making: each action has to be taken envisioned in a series of non independent actions. Reinforcement learning (RL), a field of machine learning, is an ensemble of methods based on the problem of an agent who wants to find an optimal policy, i.e. which action to choose at each time step, in order to maximize in expectation a sum of rewards in a stochastic environment (Sutton and Barto, 2018). RL is a natural tool for crop-management decision-making.
To date, RL has been poorly applied for crop-management decision-support. The most recent paper came from (Garcia, 1999), 20 years ago. Since then, a new context appeared: increasing computational power, new tools for practitioners and a new data context. The new data context comes from remote sensing, field sensors, mobile networks, drones, even social networks. It provides cheaper and more massive amounts of data that were not available at the time of (Garcia, 1999).
In DC4SCM, RL for crop-management is envisioned in a continual learning setting, learning from farmers own experiences. Such a setup requires real-world data in order to train an RL agent as a proof-of-concept of this approach. Our problem is that such RL oriented data (sequences of actions and associated plot evolution) is not available. In order to address that problem, three crops have been selected as they are cultivated both in India and Africa, in order to be relevant for both contexts (e.g. maize, pigeon pea, groundnut). We want to precisely monitor the fields and their reactions to agricultural operations all along crop cycle.
- (Sebillotte, 1974) Frédérick Garcia. “Use of reinforcement learning and simulation to optimize wheat crop technical management.” In: Proceedings of the International Congress on Modelling and Simulation (MODSIM’99) Hamilton, New-Zealand. 1999, pp. 801–806.
- (Sebillotte, 1974) Michel Sebillotte. “Agronomie et agriculture. Essai d’analyse des tâches de l’agronome.” In: Cahiers Orstom, série biologie 24 (1974), pp. 3–25.
- (Sutton and Barto, 2018) R. Sutton and A. Barto. Reinforcement Learning: an Introduction. 2nd. MIT Press, 2018.
