| With the development of 5G technology and the rapid growth of communication demand in the development of various industries in society,spectrum resources are increasingly scarce,and the limited frequency band of traditional radio frequency communication is difficult to meet people’s communication needs.In the next generation communication technology specification,visible light communication technology is proposed as a new communication paradigm.Its broad spectrum and ultra-high achievable communication rate provide new ideas for the development of communication technology.A large amount of research has proven the feasibility of visible light communication applications,but the limited coverage of visible light signals will inevitably lead to more dense deployment of access points and form a super dense visible light communication network with a large number of users.The complex and huge network structure poses significant challenges to the resource allocation of communication networks.At the same time,with the development of artificial intelligence technology,the application of machine learning in various industries is constantly emerging.Its high adaptability,strong perception and decision-making abilities have also brought new insights to the resource allocation problem of communication networks.In this paper,we will explore the application of machine learning based methods in resource allocation problems in ultra dense visible light communication networks.Firstly,we will discuss the communication architecture of visible light networks,including the distribution characteristics of visible light access points and users,and analyze the visible light channel model.A clustering scheme is proposed for the complex network structure of visible light communication,elucidating the advantages of clustering management for access points,namely reducing interference and reducing the complexity of resource allocation in the later stage.Three different clustering methods are used for experiments to verify the adaptability of the K-Means algorithm in machine learning on this issue,and adjustments are made to the K-Means algorithm to enable it to independently select the initial clustering center,added its adaptability.After the visible light access point is clustered,we will allocate the frequency resources in the cluster.In this problem,we will use spectral clustering and graph coloring to group users according to the two methods in graph theory,divide users with small mutual interference into a group,and then allocate orthogonal resource blocks for each group of users to achieve the purpose of reducing co-frequency interference.Experimental results show that spectral clustering can take into account the cumulative interference of the system,and the effect of frequency resource allocation is better than that of graph coloring and random frequency resource allocation.After achieving the allocation of frequency domain resources within the cluster,we hope to schedule the transmission power allocated to each channel of the access point reasonably,and optimize performance indicators such as system throughput and energy efficiency by reducing the transmission power of interference signals and increasing the transmission power of useful signals.In this regard,we adopt deep reinforcement learning methods DQN and DDPG for power allocation at access points,and adjust the parameters and structures of the two algorithms to address the rapid changes in visible light channel states.The experimental results show that the two algorithms have significant advantages over traditional heuristic algorithms in power allocation issues,and DDPG is slightly better than the DQN algorithm. |
