Research On Task Offloading And Resource Allocation Algorithm Of Edge Computing In Aerospace Network

The edge computing system based on the aerospace network can cope with a large number of sudden computing needs with strict delay restrictions in various remote areas in emergency disaster scenarios due to its advantages of global coverage of the aerospace infrastructure and flexible deployment of the aerospace infrastructure.In this scenario,efficient and intelligent task offloading and computing resource allocation are crucial for computing requirements with strict delay constraints in emergency disaster relief scenarios.However,the traditional task unloading and resource allocation algorithms based on multi-agent reinforcement learning train independent agents for each edge computing node,which will make the scene unable to bear the burden of agent training.The reinforcement learning algorithm that controls the scale of independent training agents while preserving the personalized decision-making characteristics of multi-agent systems based on their environment is of great research significance.To address the above issues,this article proposes a task offloading and resource allocation algorithm based on multi type agent reinforcement learning.Multiple task processing agents are set up for different tasks,and the idea of multi type agent partitioning can simultaneously consider the computational complexity and decisionmaking performance of the algorithm.The algorithm also uses a time-based task offloading mechanism to simulate the task generation process more realistically.And propose a technique of "reward group" to solve the problem of large and unstable amount of reward observation information corresponding to each agent encountered in the scenario under this mechanism.According to the simulation experiment results,in high load scenarios with smaller drone clusters,compared to the strategies obtained using single agent reinforcement learning algorithms,the algorithm proposed in this paper can achieve a performance improvement of 13.3% based on the overall system cost indicator.Compared to reinforcement learning algorithms based on a single type of multi-agent,the training speed of the algorithm in this paper has been significantly improved.In addition,based on the idea of edge computing load balancing optimization,this paper optimizes the multi type agent reinforcement learning task unloading and resource allocation decision algorithm proposed in this paper for two scenarios,namely,uniform and non-uniform distribution of ground users.We have proposed reinforcement learning performance enhancement algorithms based on dynamic partitioning of multi type agents,as well as strong chemical habit enhancement algorithms for multi type agents based on task multi hop offloading.The first algorithm dynamically classifies intelligent agents based on the scale of the drone cluster,and then dynamically selects corresponding intelligent agents for task processing decisions based on the real-time load balancing index of the drone.The second algorithm uses a DQN assisted training technique for unloading decision agents to solve the training adaptation problem of unloading decision agents in the case of delayed harvest rewards.According to the simulation experiment results,in the scenario where the unloading task generation locations are evenly distributed,the dynamic partitioning multi type agent reinforcement learning algorithm proposed in this paper can achieve a 9% overall system cost indicator performance improvement.In the scenario of non-uniform distribution of unloading task generation locations,the dynamic partitioning multi type agent reinforcement learning algorithm proposed in this article can also bring a 6% overall system cost indicator performance improvement,and the multi hop unloading multi type agent reinforcement learning algorithm can bring a 13% overall system cost indicator performance improvement.