Reinforcement learning-driven decision support for target-oriented branch pruning on urban trees

Purpose The conventional design and management of urban trees often overlook the benefits of specific canopy shapes, despite their crucial role in enhancing thermal comfort and optimizing direct sunlight utilization. This study presents a novel workflow in which designers define target leaf areas, and a decision-support algorithm guides tree management specialists in regulating growth through branch pruning to meet these targets. Design/methodology/approach We developed a framework that integrates a tree growth simulation game with a deep reinforcement learning (DRL) network for decision-making. The simulation predicts growth responses to pruning and assesses how closely the resulting structure matches the target leaf area. Based on the current tree state and reward feedback, the DRL network issues pruning decisions. The DRL network learns to optimize pruning strategies by iteratively interacting with the simulation game. Findings The configured network proved effective in navigating the complex and extensive hybrid decision space associated with tree pruning. It successfully acquired techniques to minimize penalties and consistently achieve relatively high reward scores in the game. Research limitations/implications High computational resource consumption remains a significant challenge. Additionally, the reward function lacks clear definitions that consistently guide the model toward the intended design targets. Originality/value This work establishes a novel technical pathway for implementing the proposed workflow, employing a voxel approach in the design and management of urban trees. It facilitates multifunctional tree use aligned with explicitly defined design objectives.