Enhancing Soybean Fertilization Optimization with Prioritized Experience Replay and Noisy Networks in Deep Q-Networks

This study focuses on the optimization of reinforcement learning in the Deep Q Network algorithm. This is achieved using the prioritized experience replay algorithm and Noisy Network optimization. The main goal is to optimize fertilization so that it can adapt to its environment and avoid over-fertilization. This study uses the prioritized experience replay algorithm and Noisy Network optimization to create an agent in RL that is able to explore and exploit optimally so that it can improve the precision of fertilization in soybeans. This methodology includes several steps, including data preparation, creating an environment that matches real-world conditions, and validating changes in soil nutrient conditions. The RL model was trained with PER and NN, with performance evaluated using cumulative reward, convergence speed, action distribution, and Mean Squared Error (MSE). The main results of the study show that DQN-PER NN achieves the highest cumulative reward, approaching 600,000 in 1000 episodes, outperforming standard DQN, A2C, and PPO. It also converges faster at episode 230, indicating superior adaptability. In addition, the results of this study indicate that the model that has been created is able to recommend a dose of SP36 fertilizer of 150 kg/ha, urea fertilizer of 100 kg/ha, and KCL fertilizer of 125 kg/ha. Compared with the A2C and PPO methods, the dose of urea fertilizer is reduced by 14%, KCL fertilizer is reduced by 33%, while for SP36 the difference is 23%. In Conclusion this model effectively distributes actions based on environmental conditions, which supports sustainable agriculture. In conclusion, the integration of PER and NN into DQN significantly improves exploration and decision making, and optimizes soybean fertilization. This model not only improves harvest efficiency but also encourages sustainable agricultural practices.