Abstract: During the practical operation of cartographic generalization, the continuous reduction of map scale leads to a sharp increase in the spatial density of various point annotations on maps, which easily induces a large number of spatial conflicts such as mutual overlapping and occlusion between adjacent annotations. These conflicts not only directly obscure the expression of core geographic features, damage the integrity, neatness and readability of map spatial layout, but also seriously restrict the accurate transmission, efficient reading and intuitive visualization of geographic information, thus significantly reducing the visualization effect, information transmission efficiency and practical application value of maps. Objectives： Aiming at the prominent challenges existing in point annotation overlap conflict resolution, including insufficient adaptability to dynamic environments, obvious performance degradation in high-density annotation scenarios, weak global perception ability and limited multi-constraint collaborative optimization capability, a novel approach for point annotation overlap conflict resolution based on multi-agent deep reinforcement learning is proposed. The approach is designed to make up for the inherent defects of traditional conflict processing methods in global optimization, parallel decision-making and autonomous learning, break through the bottlenecks of long computation time, unstable optimization effect and tendency to fall into local optimum in large-scale data processing, and provide a stable, intelligent and efficient technical solution for multi-scale cartographic generalization and large-scale high-density point annotation layout tasks. Methods： According to the spatial distribution characteristics, topological relations and conflict scales of map point feature annotations, agent screening and matching rules with strong adaptability are constructed to accurately identify pairwise conflicts, three-annotation conflicts and large-scale multi-annotation conflicts. Experimental data are divided into static conflict datasets and dynamic conflict datasets to build a stable, controllable and highly convergent iterative computing environment for the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm, so as to ensure stable and efficient model training. In strict accordance with cartographic specifications and geographic spatial constraints, a multi-dimensional reward mechanism including map boundary constraints, annotation-to-annotation secondary conflict constraints, annotation-to-point feature occlusion constraints and movement buffer constraints is established. Positive and negative reward combination is adopted to accurately guide agents to complete autonomous iterative learning and policy optimization. Relying on the collaborative decision-making advantages of the Actor network and Critic network in the MADDPG model, the optimal spatial positions of each annotation agent are continuously predicted, globally evaluated and iteratively optimized, so as to achieve comprehensive, efficient and stable resolution of large-scale point annotation spatial overlap conflicts without obvious secondary conflicts. Results： Comparative experimental results fully verify the superior comprehensive performance of the proposed method. Compared with traditional map annotation conflict processing algorithms and unimproved MADDPG methods, the proposed method possesses stronger operational feasibility, environmental robustness and generalization ability, with higher conflict resolution accuracy and more stable layout optimization effect. In static conflict scenarios, the method can completely eliminate point annotation overlap without generating any secondary conflicts; in complex dynamic and high-density scenarios, it can significantly optimize annotation spatial distribution, greatly improve the regularity, uniformity and rationality of the overall map layout, and effectively enhance map readability. Meanwhile, the method presents outstanding advantages in overall computational efficiency, which greatly shortens the processing time and successfully breaks through the technical bottlenecks of traditional methods such as long computing time, poor optimization effect, easy to fall into local optimum and unstable performance when processing large-scale and high-density annotation conflict data. Conclusions： The proposed method effectively overcomes many shortcomings of traditional algorithms in large-scale map point annotation spatial conflict processing, such as low efficiency, poor effect and weak adaptability. It provides solid and reliable technical support for the intelligent and efficient resolution of spatial overlap conflicts of massive point feature annotations, and can be well adapted to practical application scenarios including multi-scale cartographic generalization, dynamic map updating, high-density POI visualization and automated map production. In addition, it enriches the research system of intelligent spatial conflict processing in the geographic information field, and provides an innovative, feasible and extensible technical paradigm and reference for the research of automatic cartographic optimization, intelligent map layout and spatial information intelligent expression.