Deep reinforcement learning for spatial resource allocation: A case study of school districting

Spatial resource allocation is a multi-objective spatial optimization problem with multiple constraints. The division of school districts is a classic problem of spatial resource allocation. This paper proposes a new dynamically districting optimization method based on deep reinforcement learning to optimize the global effect of school districting. In the proposed method, the school district’s constantly adjusted allocation process is regarded as a multi-step Markov decision-making process. The method combines the advantages of a deep convolutional neural network with reinforcement learning for real-time response and flexibility, and directly learns behavioural policies based on the input of changing school district states. According to various constraints, this algorithm optimizes the distance of students to school and the utilization rate of schools, and it proposes a better allocation plan. To demonstrate its validity, the proposed method was evaluated using real datasets of two school districts in the United States. The experimental results studied in six different scenarios show that, compared with traditional algorithms, the new proposed method requires less prior knowledge and is globally optimal, and can provide a better allocation plan for school districting, which reduces the distance between students and schools and balances the utilization rate of schools.