LossLeaP: Learning to Predict for Intent-Based Networking

Abstract

Intent-Based Networking mandates that high-levelhuman-understandable intents are automatically interpreted andimplemented by network management entities. As a key partin this process, it is required that network orchestrators acti-vate the correct automated decision model to meet the intentobjective. In anticipatory networking tasks, this requirementmaps to identifying and deploying a tailored prediction modelthat can produce a forecast aligned with the specific –andtypically complex– network management goal expressed by theoriginal intent. Current forecasting models for network demandsor network management optimize generic, non-flexible, andmanually designed objectives, hence do not fulfil the needsof anticipatory Intent-Based Networking. To close this gap,we proposeLossLeaP, a novel forecasting model that canautonomously learn the relationship between the prediction andthe target management objective, steering the former to minimizethe latter. To this end,LossLeaPadopts an original deeplearning architecture that advances current efforts in automatedmachine learning, towards a spontaneous design of loss func-tions for regression tasks. Extensive experiments in controlledenvironments and in practical application case studies prove thatLossLeaPoutperforms a wide range of benchmarks, includingstate-of-the-art solutions for network capacity forecasting.