In recent years, the technology related to providing computing resources to ground users using Unmanned Aerial Vehicle (UAV) with Mobile Edge Computing (MEC) servers has received widespread attention. In edge computing scenarios with limited communication infrastructure, for the problem of difficulty in deploying conventional terrestrial base stations, this paper designs a UAV-based edge computing network (UAV-MEC) model. The network model adopts a partial offloading strategy, which enables the computational tasks of the ground devices (LD) to be partially executed on the local devices and partially offloaded to the edge computing servers of the auxiliary UAV. When facing computationally intensive and delay-sensitive computational tasks, LD and UAV have limited computational resources and energy, and how to optimize the offloading strategy becomes a key issue. In this paper, we propose an offloading optimization algorithm based on GPER-DDPG. The algorithm introduces the prioritized experience replay technique, adaptive action policy noise, and Actor network delay updating technique to improve the stability and convergence performance of the DDPG algorithm in complex environments. Experimental results show that the performance of the improved algorithm proposed in this paper improves about 10-12% over the performance of the DDPG algorithm.