Research On Network Slice Resource Allocation Based On Deep Learning

With the rapid development of the mobile Internet,the continuous growth of user service types and the massive connections of mobile terminal devices,the mobile communication network has to face higher requirements.In order to meet the diversity of service types and the differences in service requirements,network slicing technology emerged,and has become a key technology for 5G network.In this paper,the resource allocation of network slicing on the side of Radio Access Network(RAN)is studied,and the research content is mainly divided into the following two points:(1)Network slicing resource allocation based on prediction.Traditional forecasting methods always focus on accurate forecasting of time series by punishing both positive and negative errors equally.These method is not very suitable for resource allocation scenarios,because suppliers often hope that the resources allocated for network slices can carry all requirements from users as far as possible during resource management.Because the insufficient supply of resources means that the Service Level Agreement(SLA)signed with the tenants will be violated and which will result in highly churn rate of users,these all bring huge costs.Therefore,this paper proposes a novel prediction method which forecasting the upper bound of resource requirements based on the Long Short-Term Memory(LSTM)model,which reduced the degree of undersupply at resource allocation by reducing the number of cases where the predicted value is less than the true value.Aiming at the problem of limited wireless spectrum resources,the Genetic Algorithm(GA)is further used to solve the problem of constrained discrete resource allocation problem by minimizing the cost of suppliers.Finally,the virtual radio access network slice dynamic resource management system is built to carry out the experimental simulation of the scheme.the proposed scheme is simulated based on the network slicing resource management system.The results show that compared with other benchmark algorithms,the proposed scheme is more available in carrying the user's requirements and more effective in minimizing the cost of suppliers due to unreasonable resource allocation.(2)Network slicing resource allocation based on reinforcement learning.In practice,there is often a phenomenon that the user's service request mode changes with the user's service experience.However,the user's service experience is not only related to the network environment,but also largely affected by the number of resources allocated to the user.Therefore,the user's service request mode should not simply obey some independent distribution.Considering the complexity and randomness of the wireless network environment,as well as the Markov nature of the user request pattern,deep reinforcement learning is adopted in this paper to allocate resources for network slices.The proposed method aims to improve the Quality of Experience(QoE)of users and Spectral Efficiency(SE)of the system,and extends the traditional algorithm.In detail,the Deep Deterministic Policy Gradient(DDPG)framework is used for training to solve the problem of excessive action space.The Recurrent Neural Network(RNN)is introduced to deal with the time series states to fit the user's service request patterns.In order to solve the limitation of wireless spectrum resources,the Softmax layer is added to normalize the output of DDPG.Since the output actions of DDPG are continuous variables,the Wolpertinger strategy is also used to find the optimal and feasible discrete action.The proposed scheme is also verified based on the virtual radio network slicing resource management system.The experimental results show that the proposed scheme can effectively track the pattern of user request,and can ensure high QoE of users and SE of the whole system.