Optimization Design Of Mobile User Federated Learning With Data Characteristic And Channel Condition Adaptation

As a typical distributed learning paradigm,federated learning has become an enabling technique to implement network intelligence for the sixth generation(6G)communication systems,which can generate highperformance models through collaborative training of all the edge nodes,and the data privacy can be effectively ensured.However,due to the high dynamics of user behaviors,the collected data on edge nodes is usually non-independent and identically distributed(non-IID).The distribution divergence will lead to serious deterioration and degradation of the convergence rate and the accuracy performance of federated learning models.To address this issue,the existing works have proposed various solutions from both the data-enhanced and the model-enhanced perspectives,but it is still challenging to balance between the model performance and the communication and computation overheads.In addition,it is important to devise the strategy of applying federated learning algorithms to non-ideal channel transmission scenarios and dealing with the transmission failure of the model update results fed back by wireless users,which can promote efficient utilization of wireless resources.Therefore,this thesis focuses on the wireless users with non-IID data,and conducts research on designing efficient federated learning optimization methods from the perspective of data characteristic and channel condition adaptation.The main works and innovations of this thesis can be summarized as follows:First,in this thesis,a method for wireless user federated learning with non-IID data characteristic adaptation is designed,and a wireless user federated learning framework with model ensemble is proposed.In model training phase,it can effectively mitigate the impact of distribution dive.rgence by formulating independent user clusters and implementing cluster-based federated learning mechanism.In model inference phase,the cluster models are combined based on the similarity with the target tasks through model ensemble to adapt to the non-IID data characteristics and thus,the generalization performance of ensemble model can be improved.Moreover,a theoretical analysis on the model inference performance of the proposed framework is provided,which has proved that an upper bound of inference error on the test set can be established between the ensemble model and the ideal model,and effectively characterized the impact of intra-cluster and inter-cluster distribution divergence on model inference performance.Additonally,in order to design a computation-efficient user clustering method with non-IID data characteristic adaptation,the wireless user clustering problem is modelled as a coalition formation game.A joint optimization algorithm based on individual utility preferences is proposed to keep a balance between the model performance and the communication and computation overheads,which has been proved to achieve a Nashstable partition result.The simulation results on public data sets have been provided with a 10 percent test accuracy improvement in average compared to the baseline scheme,which can verify the significant performance gains with respect to the proposed method with data characteristic adaptation of wireless user federated learning on non-IID data.Second,in order to effectively deploy federated learning optimization algorithms in non-ideal wireless circumstances,a method for wireless user federated learning with non-ideal model update feedback adaptation is designed in this thesis,which can reduce the impact of transmission failure on the training efficiency and accuracy performance of federated learning models based on channel condition information.By theoretically modeling non-ideal model update feedback of wireless users,the expression of the outage probability of model update feedback is systematically analyzed.Moreover,a gradient aggregation correction method based on the non-ideal model update feedback information is proposed in this thesis,which employs the model update feedback outage probabilities to adjust the weights of wireless users in gradient aggregation and thus,the impact of channel condition differences between wireless users on the training efficiency of federated learning models can be avoided.Furthermore,it has been proved that the proposed gradient aggregation correction method can ensure that the global model update direction is unbiased,i.e.,there will be no deviation due to transmission failure.In addition,in order to effectively characterize the impact of non-ideal channel conditions on user clustering results,a wireless user clustering optimization method is proposed based on feedback channel condition adaptation to improve the adaptability and flexibility in wireless networks.The simulation results have shown that the proposed optimization method can significantly improve the test accuracy of federated learning models under different transmission conditions,and can effectively overcome the influence of transmission failure compared with other user clustering algorithms by adapting to non-ideal channel condition.