Learning to Learn How to Manage Network Resources with Loss Function Meta-Learning

The evolution of communication networks towards self-configuring systems requires the development of anticipatory approaches for network management to realize the envisioned concept of a zero-touch network orchestration. Current anticipatory network intelligence solutions rely on a well-defined loss function, which means that they require perfect knowledge of the relationship between the proactive management decisions and the consequent system performance. However, in anticipatory networking, there exist many tasks where characterizing such a relationship in advance is not possible. In such tasks, it is possible to measure the resulting performance of a management decision a posteriori, but we cannot know a priori the resulting performance of a certain management decision. A simple example of such tasks could be the maximization of the monetary profit when allocating resources to end users: when taking a certain resource allocation decision, it is possible to measure the profit afterwards, but it would be extremely difficult to determine a priori the resulting monetary profit. To close this gap, we present a novel two-fold learning approach, which is able to jointly learn the relationship between the prediction and the target management objective at the same time as it apprehends to anticipate the corresponding task. This method lays the foundations to the automated adaptation of network intelligence to specific complex objectives in zero-touch network management. We apply this method to different use cases of interest including monetary profit maximization.