The aim of this project is to leverage machine learning techniques to perform accurate user-level traffic prediction to improve theperformance of radio resource allocation. We will develop a specificmachine-learning framework that, taking as input the data about traffic flows seen from given vantage points (e.g., a base stations or the mobile core network), is able to produce as output the parameters that are needed to steer the corresponding network resource allocation mechanisms. For instance, scheduling can benefit from knowing in advance the elasticity andlatency requirements of a flow.
Our machine learning framework will consider two distinct types of parameters: (i) those related to the requirements of the different flows, such as elasticity, latency requirements and reliability, and (ii) those related to the traffic generated, such as the average and peak rates, burstiness level, etc. Note that for the latter parameters related to traffic forecasting, we can verify a posteriori the accuracy of our prediction, and this information can be used to further improve the accuracy our algorithm during its execution. In contrast, this is not possible for the flow requirement parameters, and thus in this case the learning process will be limited to the offline training phase where labeled ground truth information is available. When devising our machine learning framework, we will evaluate the performance of different machine learning techniques, paying particular attention to Deep Learning which provides a number of advantages: (i) recent improvements in computational capacity enable a more widespread use of this tool, (ii) it has been shown to outperform traditional machine-learning approaches, and (iii) in a setting with a very rich set of input parameters like the one considered in this project, this technique has the potential to significantly improve the resulting accuracy as compared to other methods.
Our objective istoforecastthe capacity needed to accommodatefuture traffic demandsin order to anticipate the minimum capacity constraint for the given traffic type. For example, while forenhanced mobile broadband (eMBB)the goal is to perform estimations of the capacity to minimize SLA violations on the on hand and prevent overprovisioning on the other hand,for ultra-reliable low-latency communications (URLLC)traffic the objective is to prevent outages completely,for example by reserving additional resources with respect to the predicted demands.For URLLC traffic, our objective is to provide a short prediction horizon, in the order of tens milliseconds.
The project will design a machine learning framework and apply it to optimize network resource allocation, and specifically scheduling. When instantiating the algorithms required for such use a case, we will take into account the following considerations: (i) the specific parameters needed for the use case, which are provided by means of machine learning, (ii) the maximum complexity allowed, (iii) the required level of accuracy, (iv) the time at which we need to take a decision, and (v) whether the decision can be easily re-evaluated later on. We will simulate the entire system including our machine learning algorithms coupled with state-of-the-art scheduler.