In the era of information explosion, the effective configuration of labels on maps is crucial for the rapid comprehension of information. The point-feature label placement problem, particularly in large-scale and high-density scenarios with spatial mutual-exclusion constraints, is a classic NP-hard discrete optimization challenge. Existing metaheuristic algorithms (e.g., Simulated Annealing and Genetic Algorithm) often struggle to achieve high-quality global layouts due to their propensity to become trapped in local optima, inefficient random point-selection processes, and inadequate modeling of the spatial mutual-exclusion and blocking constraints between labels. To address these limitations, this paper proposes a Progressive Reinforcement Learning (PRL) algorithm specifically tailored for the point-feature label placement problem. The algorithm models the label placement process as a sequential decision-making problem within the Reinforcement Learning framework, optimized through agent–environment interaction. Its core design comprises the following: (1) a staircase-like policy learning mechanism that shifts from “broad exploration in the early stage to precise exploitation in the later stage” to balance global search and local optimization; (2) a data mining-based Intelligent Action Screening (IAS) mechanism, which dynamically identifies and prioritizes “high-value action points” critical for improving layout quality by constructing the “Contribution Decline Degree” and “Contribution Support Degree” metrics. Experiments on large-scale real-world POI datasets (10,000, 20,000, and 32,312 points) demonstrate that the proposed algorithm significantly outperforms 13 state-of-the-art comparative algorithms, including Simulated Annealing, Genetic Algorithm, Differential Evolution, POPMUSIC, and DBSCAN, in terms of both placement quality and the number of successfully placed labels. It exhibits remarkable adaptability and competitiveness in handling high-density and complex scenarios.
Cao et al. (Thu,) studied this question.