News and events
22 February 2019
Research Seminar: Mobile traffic forecasting and backhaul utility optimisation using deep learning
Presented By Paul Patras (University of Edinburgh)
- N1.12 Haslegrave Building
About this event
Accurate prediction of future mobile traffic consumption is increasingly important for demand-aware resource allocation. Long-term traffic forecasting is however challenging due to the intricate spatio-temporal fluctuations associated with user movement, the substantial cost of the necessary measurement infrastructure, and the amount of post-processing involved. At the same time, mobile services continue to diversify and often have conflicting requirements, including bandwidth, sensitivity to delay, or the monetary value of the traffic generated. This poses further challenges to operators who seek to allocate resources dynamically, so as to optimise the overall network utility.
In this talk I will make the case for deep learning driven algorithms tailored to the mobile networking domain, in order to surmount such complex problems. I will first present a Spatio-Temporal neural Network (STN) that harnesses the exceptional feature extraction abilities of deep learning to achieve precise network-wide mobile traffic forecasting. By experimenting with real-world data sets, retraining this structure with prior predictions and combining its output with historical statistics, I will show that this enhanced approach can perform long-term traffic forecasting using only limited ground-truth observations, achieving 61% smaller prediction errors as compared to widely used forecasting tools. I will then introduce a general rate utility framework for logically 'sliced' backhaul networks, which encompasses all known types of service utilities, i.e. logarithmic, sigmoid, polynomial, and linear. To tackle the complexity of optimising non-convex objective functions built upon arbitrary combinations of such utilities, I will reveal a deep learning approach that can learn correlations between traffic demands and achievable optimal rate assignments. I will show that this technique can be trained within minutes, following which it computes rate allocations that match those obtained with a state-of-the-art global optimisation algorithm, yet orders of magnitude faster, while attaining 62% utility gains over a baseline greedy approach.
Paul Patras is a Lecturer and Chancellor's Fellow in the School of Informatics at the University of Edinburgh, where he leads the Internet of Things Research Programme. He received his Ph.D. from University Carlos III of Madrid and held visiting research positions at the University of Brescia, Northeastern University, TU Darmstadt, and Rice University. His research interests include performance optimisation in wireless and mobile networks, applied machine learning, mobile traffic analytics, security and privacy, prototyping and test beds.