Deep Neural Network Model for Dynamic Functional Split Management in Beyond 5G

Radio Access Network (RAN) disaggregation consists in splitting the baseband functionalities of a base station between centralized processing centers and distributed locations close to the radio units. This is becoming a key trend in the roadmap from fifth to sixth generation (6G) systems because it can lead to a more efficient utilization of computational, storage and power resources. To achieve these improvements, a proper functional split selection is required, balancing the benefits of centralization with the costs of transporting data over the fronthaul (FH). In this context, this paper proposes a Deep Neural Network (DNN)-based solution to select the functional split on a per-cell basis with the objective of minimizing energy consumption. In order to find an optimal model, a genetic-based optimization procedure is used to adjust the weights and biases of the DNN given the non-differentiable nature of the loss function. Lastly, this work analyzes the optimality of DNN output split combinations compared to an exhaustive search approach under given problem constraints.