LSDDPG: A Deep Reinforcement Learning Algorithm for UAV-assisted Task Offloading with Dynamic Feature Continuous Learning

Task offloading can significantly enhance task processing efficiency and reduce latency. However, current research often relies on fixed offloading devices, which struggle to respond quickly to sudden events or changing demands in dynamic scenarios such as disaster relief or emergency response. The mobility of Unmanned Aerial Vehicles (UAVs) enables rapid deployment to various locations to assist with task offloading; however, limitations in battery life and computational power, along with their inherently flexible nature, complicate the effective formulation of offloading strategies in dynamic environments, resulting in persistent latency issues. To address this challenge, we propose an improved algorithm–an enhanced method combining Long Short-Term Memory (LSTM) networks with Deep Deterministic Policy Gradient (DDPG) (LSDDPG). This algorithm achieves real-time optimization of task offloading decisions by continuously learning the dynamic features of system states and action variations. Specifically, LSDDPG generates superior action strategies based on the current state and historical data, learning from past states and actions to provide a more objective basis for decision-making. This approach effectively alleviates the constraints of UAVs in terms of power, computational capability, and mobility, significantly reducing task processing latency. Experimental results demonstrate that our proposed LSDDPG algorithm outperforms existing solutions in terms of task offloading latency, highlighting its potential for application in dynamic environments.