Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Machine learning for 5G and beyond

1,120 views

Published on

This presentation discusses matters of machine learning for 5G and beyond, towards a reliable and efficient reconstruction of radio maps. This presentation was given during the ITU-T workshop on Machine Learning for 5G and beyond, held at ITU HQ in Geneva, Switzerland on 29 Jan 18. More information on the workshop can be found here: https://www.itu.int/en/ITU-T/Workshops-and-Seminars/20180129/Pages/default.aspx

Published in: Technology
  • Be the first to comment

Machine learning for 5G and beyond

  1. 1. Machine Learning for 5G and Beyond: Towards Reliable and Efficient Reconstruction of Radio Maps Slawomir Stanczak Joint work with R.L.G. Cavalcante
  2. 2. Renato Cavalcante & Slawomir Stanczak • A radio map is an (unknown) function that relates a geographic location to some radio system parameter (e.g. path-loss, capacity, QoS etc). • Goal: Reconstruction of radio maps in an online fashion from user measurements 2D view: Introduction
  3. 3. Renato Cavalcante & Slawomir Stanczak 3 What are the challenges & opportunities? • High mobility è changes in network topology è wireless links exhibit ephemeral and dynamic nature è Non-stationarity (and non-ergodicity) • Noisy capacity-limited transmission exposed to interference è wireless channel is error-prone and highly unreliable • Stringent requirements of many 5G applications • Data is distributed at different locations • Models, context information and expert knowledge are available • There is a lot of structure in the channel, signals and functions • Low-complexity, low-latency implementation
  4. 4. Renato Cavalcante & Slawomir Stanczak ML for Reconstruction of Capacity Maps • Adaptive learning of long-term capacity maps
  5. 5. Renato Cavalcante & Slawomir Stanczak Madrid Scenario Madrid grid environmental model • Model of a city layed out on a grid • Raytracing data from the METIS project • Three different heights (floors) • Wrap-around model to remove edge effects
  6. 6. Renato Cavalcante & Slawomir Stanczak • Pathloss map: adaptive projected sparse-aware multi-kernel approach Kasparick M., R. L. G. Cavalcante, S. Valentin, S. Stanczak, and M. Yukawa, "Kernel-Based Adaptive Online Reconstruction of Coverage Maps with Side Information," IEEE Transactions on Vehicular Technology, vol. 65, no. 7, pp. 5461-5473, July 2016 Learning Capacity Maps: Key Ingredients
  7. 7. Renato Cavalcante & Slawomir Stanczak Learning of Pathloss Maps Each base station has access to pathloss measurements (that arrive over time) and maintains prediction of pathloss in its area of coverage Whenever new measurement arrives, the base station updates its current approximation of the unknown pathloss function Measurements may contain errors that are not uniformly distributed Requirements: • High estimation accuracy despite the lack of measurements in som e areas • Adaptivity and good tracking capabilities: Online learning based on measurements • Low Complexity: Measurements are processed in real time • Robustness: Error tolerance with respect to reported measurements
  8. 8. Renato Cavalcante & Slawomir Stanczak Learning of Pathloss Maps
  9. 9. Renato Cavalcante & Slawomir Stanczak • Pathloss map: adaptive projected sparse-aware multi-kernel approach • Traffic map: Gaussian processes, Quantile estimation, context information Learning Capacity Maps: Key Ingredients R. L. G. Cavalcante, et.al., "Toward Energy-Efficient 5G Wireless Communications Technologies: Tools for decoupling the scaling of networks from the growth of operating power," n IEEE Signal Processing Magazine, Nov. 2014.
  10. 10. Renato Cavalcante & Slawomir Stanczak Real Network: Learning Data Traffic Objective: Predict the traffic demand in a given area by using observed time series and contextual information (e.g., day of the week, holidays) • Learn from the environment • Good predictive power forecasts with confidence intervals • Do not try to learn too much!
  11. 11. Renato Cavalcante & Slawomir Stanczak Wireless Communications and Networks © Traffic Demand Pixel-wise prediction 11.06.2015 46 Above: Traffic maps for a particular time point, after training phase of 4 days Right: Aggregated traffic demand prediction for 1 day, after 4 days of training For the sake of illustration, the traffic for each test point is summed-up to give a network- wide curve Composite kernel function used Learning Data Traffic
  12. 12. Renato Cavalcante & Slawomir Stanczak • Pathloss map: adaptive projected sparse-aware multi-kernel approach • Traffic map: Gaussian processes, Quantile estimation, context information • Load estimation: hybrid-driven methods Learning Capacity Maps: Key Ingredients D. A. Awan, R. L. G. Cavalcante, and S. Stanczak, ``A robust machine learning method for cell-load approximation in wireless networks,´´ arXiv:1710.09318, 2017
  13. 13. Renato Cavalcante & Slawomir Stanczak • The rate-load mapping has a rich structure (e.g., monotonicity) that is hard to exploit in typical machine learning tools • New hybrid-driven methods: more robust and optimal, in a well-defined sense, in uncertain environments Learning Load Maps Objective: Given a power allocation for cells and the traffic demand for users, what is the load at each cell (fraction of the used resources)? Challenge: The mapping relating rate to load is highly dynamic and nonlinear owing to the interference è training must be short
  14. 14. Renato Cavalcante & Slawomir Stanczak Load estimation • We combine the tools with statistical methods to obtain probabilistic bounds • Serve as a basis for robust network optimization decisions • Example: operator is interested in knowing if particular configuration will be su cient to guarantee a certain maximum load with a given probability S. Sta´nczak November 25, 2015 (22/2 Learning Load Maps We combine the tools with statistical methods to obtain probabilistic bounds • Serve as a basis for network optimization decisions • Example: operator is interested in knowing if particular configuration will be sufficient to guarantee a certain maximum load with a given probability
  15. 15. Renato Cavalcante & Slawomir Stanczak Learning Capacity Maps
  16. 16. Renato Cavalcante & Slawomir Stanczak Traffic Forecast Optimize the Network Configuration Energy Modelling Save Energy 0 20 40 60 80 100 0 5000 10000 15000 20000 25000 30000 35000 0 12 0 12 0 12 0 12 0 12 0 12 0 12 0 Avrg.Load[%] PowerConsumption[kW] Time of Day Total Power Consumption Total Power Consumption Switch-Off Avrg. GSM Load Avrg. UMTS Load Avrg. LTE Load Energy-Saving Optimization
  17. 17. Renato Cavalcante & Slawomir Stanczak Energy-Saving Optimization # Cells # test points Considered cells per test point # opt var Time [s] Memory usage [%] #active cells after optimization 900 20.000 900 1.5 mio 276 70-80 440 900 20.000 10 0.19 mio 41 30-40 440 900 20.000 5 0.09 mio 29 30-40 516 • Simulation area: 20 km x 20 km • Number of iterations in our algorithm: 10 • Single-RAT optimization (LTE) • Using CPLEX in the iteration of the algorithm • Conventional laptop (Core i7 with 4GB of ram) Gfg
  18. 18. Renato Cavalcante & Slawomir Stanczak Outlook Reconstruction techniques based on tensors, which seem a natural fit for online tracking of channel conditions Learning A2A radio maps • D2D communication

×