Date of Award
2026
Type
Thesis
Major
Master of Science
Degree Type
Master of Science in Robotics Engineering
Department
Earth & Space Science
First Advisor
Mohammad Jafari, Ph.D.
Second Advisor
Abiye Seifu, Ph.D.
Third Advisor
Rania Hodhod, Ph.D.
Abstract
Hydroponic farming has emerged as a promising technology within controlled environment agriculture (CEA) due to its ability to increase crop productivity while reducing land and water usage. Recent advances in cyber-physical systems, digital twins, machine learning, and reinforcement learning have enabled the development of intelligent agricultural systems capable of adaptive environmental regulation.
This thesis presents a comparative study of classical and learning-based control strategies for hydroponic environmental regulation within a stochastic digital twin environment. The proposed framework models a controlled hydroponic germination-stage environment where temperature, humidity, and light intensity evolve under time-varying targets, stochastic disturbances, and synthetic sensor noise.
A stochastic multi-seed environmental simulation model was developed to approximate relative germination-stage progression under varying environmental conditions. The control problem was formulated as a continuous optimization task with the objective of maximizing simulated germination progression while minimizing energy consumption.
A classical Proportional-Integral-Derivative (PID) controller was implemented as a baseline alongside three reinforcement learning methods: Proximal Policy Optimization (PPO), Soft Actor-Critic (SAC), and Deep Q-Network (DQN).
The reinforcement learning agents were trained using the Stable-Baselines3 framework under identical environmental conditions. To evaluate robustness under uncertainty, a Monte Carlo analysis consisting of 100 independent simulation trials was conducted with randomized initial conditions and stochastic disturbances.
Experimental results demonstrate that SAC achieves the highest average reward and simulated germination performance, while PID provides the most stable and lowest-variance control behavior. Reinforcement learning controllers outperform classical PID control under stochastic conditions due to their adaptive learning capability.
This work demonstrates the effectiveness of digital twin-based reinforcement learning 2 for intelligent hydroponic environmental control and establishes a benchmark framework for future autonomous agricultural cyber-physical systems.
Recommended Citation
Massey, Andrew, "Comparative Analysis of Classical and Reinforcement Learning Controllers for Hydroponic Germination Optimization in a Stochastic Digital Twin Environment" (2026). Theses and Dissertations. 799.
https://csuepress.columbusstate.edu/theses_dissertations/799
