Towards green machine learning for resource allocation in beyond 5G RAN slicing

Abstract

The new generation of mobile communications, 5G, has been rolled out. While basic features were successively implemented, more complex ones have been left aside for a new release. Such is the case of RAN slicing, which enables the division of a radio infrastructure into software-controlled logical networks. Among the technical difficulties is the radio resource allocation since slices can be attached to contract agreements with network performance targets, and the number of resources required to support the performance varies with the signal quality of the connected devices. Machine learning solutions can predict the number of radio resources needed based on the current state of the network. Our previous work, KPI-Converter, addressed this issue with densely connected neural networks. However, amid the current global energy crisis, we should focus on more green machine learning solutions that can achieve similar performance with much lower usage of computational resources. In this work, we present KPIC-Lite, a solution for resource allocation for RAN slicing that consumes 700 to 1000 times fewer resources than our previous work while performing similarly in most tested scenarios. We introduce a new asymmetric loss function that significantly boosts convergence compared to a-OMC, the state-of-the-art loss function for operator monetary costs. Another bonus compared to a-OMC is its ability to use second-order optimisers efficiently. The second-order optimiser used in our work reduced the computational resources of the solution.