<p>Based on graph neural networks and reinforcement learning, this study designs three algorithms for multi-objective tourism route planning under different constraints: unconstrained, capacity-constrained, and demand-splitting scenarios. Furthermore, methods for quantitatively measuring these four factors are analyzed and studied. An example test was conducted on this module, and comparative and convergence tests were conducted to prove the effectiveness of the model. It can better meet the interests and needs of users. For unconstrained problems, we propose a GCN-based reinforcement learning model (GCNR) that achieves the shortest path lengths on TSP20 (Traveling Salesman Problem) and TSP50, with optimality gaps of only 0.34% and 0.35% respectively, outperforming both RLnet and PtrNet. For the multi-objective tourism path planning problem with capacity constraints, this paper designs a joint learning strategy training model that integrates supervised learning. Firstly, a model based on graph convolutional networks is trained using supervised learning to have the advantage of mining deep level feature representations. Then, a pre trained supervised learning model is used to generate high-quality initial solutions for the reinforcement during the training process and effectively improving the learning efficiency of the model in solving large-scale problems; The paper proposes an algorithm that combines heuristic search and reinforcement learning for the multi-objective tourism path planning problem with demand splitting. During the training process, heuristic algorithms are used to further optimize the solution of reinforcement learning, and the direction of model gradient descent is determined through two parts: the evaluation function of the baseline network and the improvement amplitude of the heuristic search algorithm.</p>