Research On Environment-Aware Caching In Multiple Base Stations Edge Networks Based On Attention Mechanism

With the development and progress of society,due to the continuous advancement of wireless network technology,the number of applications and the amount of data that follow has increased,and traffic has surged.This means that much high-quality content is favored by most people and multiple repeated requests occur,which reduces the efficiency of requests and also leads to a waste of resources.In the multi-layer heterogeneous edge network,the integration of caching technology and edge computing networks brings huge benefits to the entire mobile network system.Edge caching technology can efficiently reduce the resource and time costs in the redundant request and transmission process.Nevertheless,in the actual cases,the cache space is limited,and it is unrealistic to cache all of the popular content.It is necessary to find a cache replacement strategy to select and update cache content.There are still some problems in the current replacement work,traditional cache replacement algorithms are mostly static algorithms based on time and frequency statistics,some optimized algorithms rely more on content popularity and lack considering the complex network environment of heterogeneity,simultaneously,there is the phenomenon of injustice global optimization and local optimization,which paying attention to the overall benefits at the expense of the local optimum and neglecting the interaction between edge nodes in different regions.Based on the Markov process,this thesis mainly considers the cost of content request and transmission and models the cache architecture in a multi-base station environment.In order to make the edge servers have the perception ability in the multinode edge network,this thesis uses the Actor-Critic network with an attention mechanism to solve the problem and make strategic choices for replacement.The first method proposed is a neighbor-aware cache replacement algorithm named NAC based on the multi-head attention mechanism.Firstly,this research divides the information in the state space to form three feature subspaces represented by the information of the base stations,the characteristics of the cached content,and the features of the user’s historical requests.The multi-head attention layer added to the Critic network processes the three aspects of information and can observe the overall situation including the surrounding adjacent areas.The Actor network makes replacement decisions,and the Critic network based on global information evaluates the local decisions of the Actor network,to help agents realize the perception of time and space characteristics that are difficult to predict in the environment.In addition to learning-based optimization,this article also proposes an engineering solution with the idea of attention mechanism as the core,the NILFU algorithm.On the basis of the Least Frequently Used algorithm,the temporal and spatial characteristics of adjacent service areas are added.Services obtain the influence of neighbor nodes based on the historical request situation in the neighbor areas,and the arrangement of content priority is adjusted to optimize the traditional Least Frequently Used.In summary,based on the attention mechanism and edge caching service,this thesis proposes the network environment perception replacement algorithm based on multi-agent deep reinforcement learning and the least access frequency replacement algorithm based on the influence of neighbors.The proposed method makes full use of the communication between base stations to exchange adjacent information,thereby reducing the pressure on the backbone network and further improving user satisfaction.The simulation results verify the feasibility and effectiveness of the algorithm.