Deep Reinforcement Learning Based Automatic Control in Semi-Closed Greenhouse Systems

This work proposes a novel deep reinforcement learning (DRL) based control framework for greenhouse climate control. This framework utilizes a neural network to approximate state-action value estimation. The neural network is trained by adopting a Q-learning based approach for experience collection and parameter updates. Continuous action spaces are effectively handled by the proposed approach by extracting optimal actions for a given greenhouse state from the neural network approximator through stochastic gradient ascent. Analytical gradients of the state-action value estimate are not required but can be computed effectively through backpropagation. We evaluate the performance of our DRL algorithm on a semi-closed greenhouse simulation located in New York City. The obtained computational results indicate that the proposed Q-learning based DRL framework yields higher cumulative returns. They also demonstrate that the proposed control technique consumes 61% lesser energy than deep deterministic policy gradient (DDPG) method.