|PARERA SOTOLONGO CLAUDIA||Cycle: XXXII |
Tutor: RIVA CARLO GIUSEPPE
Advisor: CESANA MATTEO Major Research topic
:Enhancing Anticipatory Networks: A Transfer Learning ApproachAbstract:
Machine learning will play a major role in handling the complexity of future mobile wireless networks by enhancing network management and orchestration capabilities. Due to the large number of parameters that can be configured in the network, collecting and processing high volumes of data is often unfeasible during network runtime. This calls for taking resource management and service orchestration decisions with only a partial view of the network status.
Transfer learning is a machine learning paradigm that aims at improving the prediction performance of a learning task with limited amount of information. Motivated by this fact, in this thesis we provide a transfer learning framework that can be applied to anticipate and further adapt network decisions having a partial view of the network. Predictions can be carry out with improved performance when information in the target network domain is missed. To support this, we test our framework in three different use cases using real data sets collected from 4G commercial networks. Namely, they are: (i) Tilt-Dependent radio map predictions, (ii) Channel quality predictions and (iii) Large scale counter predictions.
The main contribution of this work is applying transfer learning to mobile communication networks. Firstly, we show the amount of knowledge to transfer and when transfer learning should be performed to avoid negative transfer. Secondly, we give guidelines to the operator about data selection strategies in the target domain. We show how to improve the performance by optimizing the model on the source domain. Finally, we provide a transfer learning framework tested on spatial-temporal data for regression.