Reinforcement Learning (RL) is an area of machine learning in which an agent continuously interacts with the environment where it operates to establish a policy - a mapping between environment states and action- that maximizes an expected reward. RL is particularly suited for decision making problems that need to trade off between long term and short term rewards and where the environment is complex and uncertain.

Optimizing crop yield in agriculture is one of such complex problems that may be addressed well with RL. Fruits take time to gain weight and ripen. The weight and timing of the harvest depend on factors such as temperature, CO2 and light level they are subjected to in the period between flowering to ripen fruits. Focusing on one week of harvest might have unexpected impact on later weeks. These are further complicated by the uncertainties in weather.

In order to help growers optimize their harvests, we explore the usage of RL for grower decision support tool which provides advise on the environmental condition settings. We experienced that developing RL application has become easier and faster because of the available open source toolkits. One such toolkit, OpenAI Gym, provides a lot of algorithms and functionalities that we leveraged. This enables us to focus more on the application development itself (e.g. designing a good reward function).

In this talk, we will discuss how we translate the yield optimization into RL problem, including the usage of Proximal Policy Optimization to design the policy and the design of the reward function. We will further discuss how we utilize OpenAI Gym and our learning from using it. Finally, we will evaluate the results and discuss further works.