Abstract: In recent years, cloud computing has been widely adopted by individuals and organizations. However, as clouds are typically deployed far from end users, processing tasks may incur considerable latency. Therefore, it is both reasonable and necessary to leverage the advantages of edge computing to develop a cloud-edge collaborative network architecture. A tree-based genetic programming hyper-heuristic algorithm (TBGP-HH) is proposed to address the task offloading problem in cloud-edge collaborative environments. The algorithm can dynamically generate service placement and task offloading strategies according to resource configurations and input tasks, thereby enhancing task processing reliability and reducing latency. First, a set of low-level heuristic algorithms is designed based on the objectives of task offloading. These heuristics are then encoded as genes to construct individual encoding trees for population initialization. Next, the population evolves iteratively through selection, crossover and mutation operations, with an elitism strategy adopted to preserve high-quality individuals. Finally, a heuristic algorithm with strong performance in both reliability and latency optimization is generated to guide service placement and task offloading. Experimental comparisons with benchmark algorithms in real-world application scenarios demonstrate that TBGP-HH consistently improves task processing reliability and reduces latency across diverse scenarios, with overall performance outperforming recent task offloading algorithms.