Research On Adaptive Equalization Technology Based On Scalable Ensemble Tree In Optical Interconnection Of Data Center

At this stage,data center optical communications are mainly based on short and medium-distance transmission.With the explosive growth of network traffic,higher requirements are placed on the data center optical interconnect transmission rate.Under the dual influence of low-cost devices in high-rate,intensity modulation/direct detection(IM/DD)systems,the transmission signal in the data center optical interconnection system will be seriously damaged,while the traditional single-function static equalization technology while the traditional single function static equalization technology can't meet the needs of dynamic compensation for multiple damages in the communication system.In the field of optical communications,neural network technology has been used to achieve signal equalization.However,the neural network is still unsatisfactory in terms of system complexity.It is an algorithm that exchanges high complexity for high accuracy.As another important branch of the machine learning algorithm group,the ensemble tree model not only has good performance but also has low complexity in the application scenarios of processing smaller data sets.In this paper,a self coding scheme is proposed around the ensemble tree algorithm,and the corresponding equalization scheme is proposed and verified for the problem of signal damage in the medium and short distance optical communication.The main research work and innovations of this paper include:Firstly,the main machine learning technology in the existing optical communication network is still neural network,and the ensemble tree model represented by XGBoost has been widely concerned in the recent data mining related competitions.The technical characteristics of XGBoost algorithm compared with neural network are analyzed,summarized and verified in the data center optical communication scene.The experimental results show that XGBoost runs fast,has unique advantages in processing structured data,and has strong interpretability.In optical communication scenario between data centers,XGBoost runs 30 times faster than neural networks,which greatly improves training efficiency of the model.Secondly,for the existing end-to-end learning mainly depends on the neural network technology,and the neural network algorithm has the problems of parameter adjustment difficulty and low model reuse rate,a learning scheme of end-to-end optical communication based on ensemble tree algorithm is studied and designed,and the feasibility of autoencoder is verified.The experimental results show that the autoencoder based on the ensemble tree algorithm can obtain the reconstructed input with MSE of 1.4699 × 10-6 in 1000 randomly generated samples,and can completely recover the input after the rounding operation.Thirdly,for the problem of signal damage in optical communication of the data center,the scheme of applying XGBoost algorithm to the short-distance optical communication equalization scenario is proposed,and the application effects of the three receiving-side equalization algorithms are compared.The experimental results show that XGBoost can get a lower bit error rate,and with the increase of distance,it can further reduce the power budget,especially in the optical communication scenario between data centers.Compared with neural network,XGBoost can achieve 2dB transmission power improvement,which verifies that XGBoost algorithm has good performance in the field of optical signal equalization based on data centers.