Multi-Agent Deep Reinforcement Learning Offloading Algorithm Based on Resource Reservation and Task Adjustment

With the popularization of the Internet of Things and mobile Internet, the data shows explosive growth. Adopting cloud computing mode will lead to network congestion and high energy consumption, and can't meet the low latency requirements of applications. Computing offloading technology offloads tasks to the edge of the network for processing, which can effectively make up for the shortcomings of the cloud computing model. However, in the multi-user and multi-server edge system, it is challenging to formulate the task offloading strategy. In this paper, we consider divisible and latency sensitive tasks and edge server load dynamics. Then we model the edge computing scenario, and formulates a task offloading problem to minimize the overall task processing latency while meeting the application latency requirements. We propose a resource reservation strategy to expand the candidate destination nodes for task offloading and propose task adjustment strategy to adjust the resource allocation of elastic tasks which can improve the task carrying capacity of MEC system and then improve the average quality of experience. We use a multi-agent deep reinforcement learning method to solve optimization problems. The simulation results show that our method can significantly reduce the application latency rate and improve the average quality of experience and resource utilization of MEC system.