Deep Reinforcement Learning for Greenhouse Climate Control
Lu Wang, Xiaofeng He, Dijun Luo
Abstract
Worldwide, the area of greenhouse production is increasing with the rapid growth of global population and demands for fresh food. However, the greenhouse industry encounters challenges to find automatic control policy. Reinforcement Learning (RL) is a powerful tool in solving the autonomous decision making problems. In this paper, we propose a novel Deep Reinforcement Learning framework for cucumber climate control. Although some machine learning methods have been proposed to address the dynamic climate control problem, these methods have two major issues. First, they only consider the current reward (e.g., the fruit weight of the cucumber). Second, previous study only considers one control variable. However, the growth of crops are impacted by multiple factors synchronously (e.g., CO2 and Temperature).To solve these challenges, we propose a Deep Reinforcement learning based climate control method, which can model future reward explicitly. We further consider the fruit weight and the cost of the planting in order to improve the cumulative fruit weight and reduce the costs.Extensive experiments are conducted on the cucumber simulator environment have shown the superior performance of our methods.