Multi-Agent Reinforcement Learning for Crop Planning Model with Interactive Multi-Stakeholder*

Current crop planning models aim to optimize planting schemes for farmers to maximize their social and ecological benefits within specified constraints, or to optimize planting structure schemes for related departments. However, farmers' planting strategy is closely tied to national planting subsidy policies, and governmental optimization and adjustment of planting structures require farmers' compliance. Focusing solely on the planting plans of either farmers or relative departments may not effectively address real-world scenarios. We propose an Interactive Multi-Stakeholder Crop Planning Model (IMSCPM) based on multi-agent reinforcement learning approaches. This model aims to provide specialized crop planning recommendations to individual farmers while simultaneously offering guiding policy suggestions to relevant regional departments. Stakeholders involved in the crop planning process are represented as multiple intelligent agents with specific objectives, allowing them to interact with each other. To address challenges posed by the proliferation of network input dimensions, the problem of strategy pursuit between two types of agents and the instability of the multi-agent environment caused by the excessive number of agents in the Regional-Farmer planting planning model, we introduce the methods of “Integrated observation”, “Asynchronous training” and “Planting consultant”. We constructed a multi-stakeholder interactive crop planning environment, and obtained strategies that are stable and consistent with the values pursued by each agent. The results demonstrate that our crop planning model based on this novel framework is effective for agricultural system with multi-stakeholders.