Performance Optimization Of Few-shot Leraning In Wireless Networks

6G will accommodate a variety of application scenarios,especially autonomous driving,mobile extended reality,smart grid,and haptics.Network intelligence and automation capabilities need to be enhanced to meet service quality requirements in different application scenarios.As a result,AI is considered a key technology in 6G,which is expected to drive the evolution of future communication paradigms and the transformation of next-generation wireless network architecture.As the computing power of network devices continues to escalate,more and more artificial intelligence as well as deep learning applications are being deployed at the network edge.However,due to limitations in data sensing capabilities,it is difficult for network edge nodes to collect large amounts of data,which makes it difficult to generate deep learning models with high performance.Few-shot learning can be utilized to solve the problem of scarce data at the network edge.where the goal is to be able to quickly generalize to new tasks containing only a small number of samples using prior knowledge.Meta-learning provides a simple and flexible solution for few-shot learning.In wireless networks,meta-learning models can be generated using historical data from base stations to provide a priori knowledge for fewshot learning at the network edge.However,due to the high uncertainty of the wireless channel environment and the large diversity of network user behavior,the single meta-learning model is difficult to meet the needs of all users.Moreover,when there is a large gap between the user target task and the meta-learning model training task,the performance of few-shot learning will decline.In addition,deploying few-shot learning at the edge of wireless network needs to consider the communication cost caused by the migration of the meta-learning model between network nodes.Therefore,this paper carries out specific research on the performance optimization method of wireless network small sample learning.The main research work can be summarized as follows:Firstly,this thesis studies a cooperative meta-learning model selection optimization method for wireless network few-shot learning.Firstly,to solve the problem that a single meta-learning model is difficult to adapt to all users’ needs,this thesis proposed a meta-learning model distribution and few-shot learning method based on multicast model distribution.Divides the meta-training data into multiple copies to generate multiple lightweight meta-learning models respectively,and selects a meta-learning model for each user based on the similarity between the characteristics of the user’s target task and the training task of the meta-learning model.Secondly,in order to balance the performance of few-shot learning and the communication cost of meta-learning model transmission,a collaborative meta-learning model selection optimization problem is proposed.Finally,the optimization problem is modeled as a coalition game problem and a joint optimization algorithm that can balance the model accuracy performance of few-shot learning and the communication cost is provided.Simulation results show that the proposed algorithm can improve model accuracy and reduce communication resource consumption significantly.Secondly,this thesis studies a meta-learning model generation optimization method for few-shot learning in wireless networks.To solve the problem that the performance of few-shot learning decreases when the training task of the meta-learning model differs greatly from the target task,this thesis proposed a few-shot learning solution based on grouped metalearning.Users are grouped according to the characteristics of the user target task and the channel condition information,and the meta-learning model is trained for each group of users.The target task information of each group of users is aggregated as the initial information of the learning model training of each group.Secondly,to improve the performance of the meta-learning model on the small sample learning tasks of users in the group,the training tasks participating in the updating of meta-learning model parameters are selected in the process of grouping meta-learning and training.Finally,in order to balance the performance of the meta-learning model on specific objective tasks and other tasks,a meta-learning model is proposed to train the task selection optimization problem,and the corresponding algorithm is designed to solve the optimization problem.Simulation results show that the proposed algorithm can effectively improve model accuracy and save communication resources.