Year of Award

2023

Document Type

Thesis

Degree Type

Master of Science (MS)

Degree Name

Computer Science

Department or School/College

Department of Computer Science

Committee Chair

Doug Brinkerhoff

Commitee Members

Ben Colman, Jordan Malof

Keywords

Reinforcement Learning, Controlled Environment Agriculture

Publisher

University of Montana

Abstract

Demands on agricultural systems will increase as the world population continues to grow. Controlled Environment Agriculture (CEA) is an old idea gaining more recent momentum due to advances in automation, lighting, and climate control. One of the fundamental challenges of CEA is optimal control of the growing environment. We present a method for optimal control in CEA utilizing Deep Reinforcement Learning (DRL). DRL comprises a set of machine learning algorithms that optimize a user-defined reward function. We define this function such that the reward is high when the growing environment is ideal for crop production, and the reward is low when the environment is not ideal. To facilitate our investigation, we created a simulated greenhouse environment defined by a set of differential equations. The critical metrics for crop production simulated by the differential equations are temperature and humidity in our simulation. The simulated environment allows the DRL algorithms to directly control the temperature and humidity. Over many simulations, the DRL algorithm "learns" how to control the greenhouse environment through trial and error. Our work shows that DRL algorithms can control temperature and humidity over long time intervals. We also compare DRL algorithms classes showing that Off-policy algorithms outperform On-Policy algorithms. Lastly, we demonstrate that Off-Policy DRL algorithms are robust to scenarios with shifting climatic distributions and changing greenhouse dimensions.

Download

Share

COinS
 

© Copyright 2023 Daithi Kahele Leilehua Martin