AI-based algorithms and experimental evaluation for beyond 5G

The fifth-generation technology standard for cellular networks (5G) aims to provide high throughput, reduced latency, massive connection and a shift from service-orientation to user-orientation in requirements and innovations. To achieve the target goals, efficient resource allocation and management are required. In this regard, the concept of Network Function Virtualization (NFV) has been introduced, as well as the one of Software Defined Network (SDN). NFV is a state-of-the-art approach which replaces whole classes of network’s nodes functions, which were generally deployed on dedicated hardware, with pure software implementations. The virtualization of radio access networks (vRAN) comes under the NFV paradigm. The goal of vRANs is to centralize the virtualized radio access point (vRAP) processing stack, and to efficiently map the users’ requirements into radio and computational resources allocation. This progress is required by mobile operators in order to support the rising traffic demand and ensure the varied quality-of-service (QoS) requested at affordable cost. But currently, due to the high complexity of the dependencies between computing and radio resources, there still are not sufficiently efficient solutions to satisfy the telecommunications industry. In this work it is investigated and presented the design of a dynamic resource controller for vRANs based on machine learning techniques. Such a design allows to continuously adapt the resource allocation to the actual demand across vRAPs. In our configuration, the first acting entity is an encoder which accomplishes the task of translating the high-dimensional state information into a simpler representation. The state information is summed up in a vector per vRAP containing samples of the amount of available data and of the average and variation of the signal-to-noise ratio experienced by the connected users. In this way the details about the traffic load, the channel quality and the users’ mobility of each RAP is gathered. It has been chosen to implement an independent deep autoencoder with nonlinear activation functions and sparsity constraint per context feature, in order to get the temporal correlation between its consecutive samples. The encoder is then followed by the proper resource manager, which has the duty of understanding the relationship between radio solutions and the corresponding computational load. The resource controller is composed by a CPU and a radio policy. The first one is responsible for the distribution of the available computational resources among all the RAPs. The last one is responsible for the choice of a bound for the modulation and coding schemes (MCS) to be adopted by each RAP, which strictly depends on the allocated computing capability. The radio policy relies on supervised learning, and more precisely on a classifier modelled with a feed-forward neural network that indicates if the decoding error probability associated to a certain MCS bound is below a given threshold or not, given the state information and the CPU time allocation. For the CPU policy, instead, it has been designed a deep deterministic policy gradient algorithm based on an actor-critic model-free solution. Both actor and critic have been deployed with a different neural network. The actor, given the state, has to estimate the CPU allocation policy, while the critics estimates the action-value function, namely the reward, based on both the state and the actor’s chosen action.