Thinking as human: Self-reflective reinforcement learning framework for fertilization decision-making

Variable-rate fertilization based on reinforcement learning has achieved significant success in simulated environments. However, issues of insecurity and inefficiency led by blind exploration mechanism constitute impediments to apply reinforcement learning in the real-world. In this paper, we proposed a self-reflective reinforcement learning framework based on bidirectional inference network for fertilization decision-making. Such framework conducts risk assessments of potential outcomes before action execution by employing bidirectional inference network, thereby mitigating the risks associated with blind exploration. The forward inference of this network, which is also manifested as predictive ability, models how crops would grow under specific fertilization strategy, while the inverse inference, seen as the reflective capability, seeks the optimal fertilization strategy to guide the crop toward the best possible growth. This study is the first research to integrate the capability of both forward state prediction and inverse action reflection within a unified architecture of invertible neural network, and accordingly developed a variable-rate fertilization algorithm, augmenting policy learning of reinforcement learning. Additionally, targeting on the variable-rate fertilization, a comprehensive profit evaluation system integrating “fertilization times” and “fertilizer-caused yield” is devised to address the existing inadequacies in profit evaluation. Compared to competitive benchmarks, the results of experiments demonstrate that our variable-rate fertilization algorithm not only achieves optimal planting outcomes but also effectively accelerates the convergence. More critically, the proposed method enables target-oriented exploration, thereby making the execution of fertilization safer and more efficient.