The Application And Optimization Of Transfer Learning For Internet Of Things

With the technical support of Artificial Intelligence(AI)algorithm,the edge of Internet of things(IOT)has gradually changed from data source to intelligent application portal.A large number of intelligent applications represented by real-time perception and dynamic decision-making are emerging at the edge of the network.But,AI based solely on cloud has some insurmountable shortcomings,such as high communication pressure,slow application response,poor data privacy and so on,Therefore,it is imperative to push AI to the edge of the network.However,the implementation of edge AI still faces challenges such as heterogeneous data,insufficient computing and storage resources and long algorithm convergence time.Through knowledge reuse,transfer learning can not only reduce the dependence on data,computing and other resources,but also improve the performance of the algorithm,so it has become an effective solution.In this paper,transfer learning is used to solve the problems that deep learning models are difficult to train at the edge of the Internet of Things and the efficiency of agent interaction is low in multiagent scenarios.It is studied from two aspects:constructing the joint optimization framework of transfer algorithm performance and edge resources,using knowledge transfer to improve the interaction efficiency of agents.The main research contents are as follows:(1)Aiming at the problems of low transfer efficiency and difficult model training in IOT edge caused by insufficient resources such as communication and computing,a multipoint to multipoint transfer learning algorithm suitable for wireless network is proposed.Based on the data diversity of heterogeneous devices at the edge of IOT,a multi-source and multi-target transfer learning algorithm is designed.Considering the grouping characteristics of multi-source and multi-objective,a wireless multicast communication model based on orthogonal frequency division multiplexing(OFDM)is constructed,on this basis,the joint optimization problem of source device selection and wireless network resource allocation is established.Considering the coupling of optimization variables in the optimization problem,a source device selection scheme based on source device data similarity and communication resource utilization is proposed,and the communication resources are allocated after the source device selection scheme is obtained.The final simulation results show that compared with randomly selected transfer data,the proposed algorithm can achieve high accuracy under the condition of limited resources,and has a great improvement in time delay and scalability.(2)Aiming at the disadvantages of low interaction efficiency and long convergence time of traditional multi-agent algorithm in solving the tasks of multi device joint decision-making and prediction under the Internet of things,a Multi-Agent Reinforcement Learning Algorithm Based on online knowledge transfer is proposed.Considering the partial observability and non-stationary environment in the multi-agent scene,the soft target with global information is generated by using the soft target mechanism,and the soft target with global information is regarded as one of the optimization objectives of the agent critical network by using the online knowledge transfer method.In order to further improve the interaction efficiency between agents,four soft target generation schemes reward based,linear combination,MinLogit and similarity clustering are proposed.Finally,the experimental results in the grounded communication environment show that the proposed algorithm can help the agent converge quickly and obtain higher reward.