Research On Resource Allocation Algorithm For Wireless Federated Learning

With the large-scale commercial use of the Fifth Generation(5G)network,the number of devices and the amount of data generated show an explosive growth trend in the wireless network.It is expected that in the future Sixth Generation(6G)network,advanced artificial intelligence(AI)will be widely used to support the rapid collection and processing of massive data,real-time monitoring of network status and network optimization.Benefiting from its advantages in reducing the data transmission load and protecting users’ privacy data,federated learning(FL)is considered to be one of the key technologies enabling future intelligent wireless communication networks.Although introducing FL into wireless networks can provide many conveniences for network in data processing,network state detection and optimization,its efficiency is always affected by resource allocation scheme because both local and global parameters are transmitted through wireless channels.How to support the effective deployment of FL in wireless networks with limited computation and communication resources while taking into account multiple indicators such as latency,energy consumption and learning performance are the main difficulties in the design of wireless FL scheme and the corresponding resource allocation algorithm.In this regard,this thesis studies the resource allocation algorithm for wireless FL.The main contributions are as follows:Firstly,considering the limited computation capabilities of low-cost devices and the difficulty of balancing multiple performance indicators in wireless FL,an edge computing-based wireless FL scheme is designed and a joint resource optimization algorithm that can balance multiple indicators is proposed.Specifically,a non-convex optimization problem with the goal of minimizing the weighted sum of the energy consumption of small base stations(SBS),the latency of wireless FL and its learning performance is formulated in this thesis.Through problem decoupling,the original problem is decoupled into three subproblems and then the corresponding optimization algorithms are given respectively.Finally,the simulation results verify the performance of the proposed algorithm in reducing the total cost of edge computing-based wireless FL.Secondly,focusing on the prominent contradiction between the requirement of training and transmitting the high-dimensional FL model and the limited computation capabilities of end devices and the limited bandwidth resources of wireless network,a network pruning-based wireless FL scheme is designed.Then,an algorithm for jointly optimizing the pruning parameters and resource allocation is proposed.Specifically,the convergence behavior of above FL scheme is firstly analyzed in this thesis,and the convergence upper bound which reflects its learning performance is obtained.After that,an optimization problem for optimizing local pruning rates and bandwidth allocation is formulated to minimize the weighted sum of single-round wireless FL latency and its convergence upper bound.Through problem decoupling,the original problem is decoupled into two subproblems and the corresponding optimal solution is obtained.Finally,the performance of the proposed algorithm is verified by simulation results.Thirdly,in view of the large communication overhead caused by the frequent decoding of all local parameters in traditional multiple access scheme-based wireless FL,and combining the advantages of network pruning in accelerating model training,an over-the-air computation and network pruning-based wireless FL scheme is designed.And a joint optimization algorithm for sensor selection,transmission power and pruning rates is proposed.Specifically,a wireless FL latency minimization problem constrained by the derived convergence upper bound of proposed scheme is formulated in this thesis.After that,an iterative optimization algorithm is proposed by simplifying the above long-term optimization problem into a single-round one and then decoupling it.Finally,the simulation results prove the effectiveness of the proposed algorithm.