Machine learning for decentralized and flying radio devices
2 likes700 views
The document discusses advancements in machine learning applications for decentralized and flying radio devices, emphasizing their robustness and efficiency in handling model uncertainties and large datasets. It explores various use cases, including modulation detection, resource allocation, and optimal placement of flying relays, highlighting the need for collaboration and coordination in decentralized wireless networks. Additionally, it introduces team deep neural networks (DNNs) for improving power control and discusses learning-based positioning through radio map reconstruction for enhanced communication in UAV-assisted networks.
![Offline path planning
-
[1] R.Gangula, D.Gesbert, et.al “Trajectory Optimization for Mobile Access Point” Accepted in Asilomar 2017.
[2]A. T. Klesh, P. T. Kabamba, and A. R. Girard, “Path planning for cooperative time-optimal information collection,” in American
Control Conference, June 2008, pp. 1991–1996.
[3] Q. Wu, Y. Zeng, and R. Zhang, “Joint Trajectory and Communication Design for UAV-Enabled Multiple Access,” , Apr. 2017.
Small/Moderate flying budget Large flying budget](https://image.slidesharecdn.com/machinelearningforflyingradio-180130153743/85/Machine-learning-for-decentralized-and-flying-radio-devices-10-320.jpg)
![Machine Learning for 3D map reconstruction
Proposed concept:
1- Learn propagation parameters using previous EM algorithm
2- Soft-classify users into LoS/NLoS
3- Reconstruct 3D map from “radio shadow” data
Connections with
-3D Imaging with WiFi [Y. Mostofi et al. @ UCSB]
-GPS based 3D imaging [Madhow et al @ UCSB]
[Esrafilian, Gesbert, Globecom 2017]](https://image.slidesharecdn.com/machinelearningforflyingradio-180130153743/85/Machine-learning-for-decentralized-and-flying-radio-devices-16-320.jpg)
![Joint 3D and radio map reconstruction
17
An orthoimagery of
an area at center
Washington DC, USA
[Chen, Esrafilian, Gesbert, Mitra, Robotics, Science and Systems, MIT, 2017]
Radio map
reconstruction
SNR/Radio map estimate
3D city map estimate
Joint approach](https://image.slidesharecdn.com/machinelearningforflyingradio-180130153743/85/Machine-learning-for-decentralized-and-flying-radio-devices-17-320.jpg)
Machine learning for decentralized and flying radio devices
- 1. Machine Learning for Decentralized and Flying Radio Devices January 29, 2018 Prof. David Gesbert EURECOM – Sophia Antipolis, France
- 2. 29/01/2018 - EURECOM GENERAL - p 2 EURECOM - Sophia Antipolis • Sophia Antipolis: One of Europe's largest technology parks • EURECOM is a ICT-oriented lab and English-speaking graduate school • Co-owned by IMT Paris, TUM, Chalmers, Politecnico Torino, CTU Prag, NTNU, Aalto, etc. (15 academic/industrial shareholders)
- 3. AI For Wireless 29/01/2018 - - p 3 Why? Robustness wrt model uncertainties Ability to make use of large measurement data sets Efficient implementations PHY applications Modulation detection Error correcting code (decoding) Beam search (hybrid massive MIMO) Etc. LINK/Network level applications Traffic classification Predictive resource allocation SDN optimization Etc. QoE enhancement User-machine interction
- 4. AI For Wireless: New perspectives 29/01/2018 - - p 4 Autonomous radio devices Dealing with measurements uncertainties Dealing with decentralized decision scenarios Case 1: Team DNN for cooperative decentralized communications Case 2: Machine learning for optimal placement of flying relays
- 5. Team Learning for decentralized wireless networks 29/01/2018 - - p 5 sharing/caching of user’s data symbols Imperfect CSI sharing x2=w2(H(2) )s x3=w3(H(3) )s x1=w1(H(1) )s Collaboration/Coordination Pilot allocation power control beam alignment resource allocation Caching Decentralized/heterogeneous setting Lack of efficient backhaul Observations are noisy, local Need for robust decision, predict decisions of others!
- 6. Introducing Team DNNs 29/01/2018 - - p 6 Goal 1: Learning to collaborate/coordinate Goal 2: Learning to deal with uncertainties -> one’s own uncertainties -> other’s uncertainties
- 7. Team DNN example: Power Control under arbitrary noisy feedback 29/01/2018 - - p 7 Decentralized power control to maximize total throughput Each TX observes gain feedback with arbitrary observation noise Each TX needs to make instantaneous power control decision (de Kerret & Gesbert2018,Arxiv)
- 8. Illustration of Team DNN behavior 29/01/2018 - - p 8 Feedback Uncertainties T-
- 9. UAV-aided wireless networks UAVs provide connectivity where and when it matters Dramatically improves propagation statistics Many possible scenarios Data scenario: broadband access vs data collection (IoT) Placement scenario: static placement vs. trajectory planning Optimization scenario: Offline vs. online (learning based) - Initial UAV position Optimal position Learning trajectory Optimal path
- 10. Offline path planning - [1] R.Gangula, D.Gesbert, et.al “Trajectory Optimization for Mobile Access Point” Accepted in Asilomar 2017. [2]A. T. Klesh, P. T. Kabamba, and A. R. Girard, “Path planning for cooperative time-optimal information collection,” in American Control Conference, June 2008, pp. 1991–1996. [3] Q. Wu, Y. Zeng, and R. Zhang, “Joint Trajectory and Communication Design for UAV-Enabled Multiple Access,” , Apr. 2017. Small/Moderate flying budget Large flying budget
- 11. Learning-based positioning Training Data Collection (RSSI, traffic,) Radio Map Reconstruction Automatic Drone Position Computation Autonomous Drone Navigation New approach: Simultaneous Learning and Positioning (SLAP) Key idea: Learn positioning through radio map reconstruction
- 12. Map-based positioning 29/01/2018 - - p 12 1) UAV Learns LoS opportunities from above, with guarantees. 2) Use limited drone training locations only (flying time is costly!) Suitable for micro-UAV (low altitude) high accuracy positioning: Top view city map UAV-user radio map End-to-end capacity map BS UE
- 13. Learning Segmented Channel Models - Received power map UAV 100 meter above center of Bristol
- 14. Kernel-based Radio Map Reconstruction Given a position x, find a set of points from training data set Estimate segment labels through: using the kernel function Soft SNR reconstruction
- 15. Radio map reconstruction Dense Training Sparse Training Role of 3D maps?
- 16. Machine Learning for 3D map reconstruction Proposed concept: 1- Learn propagation parameters using previous EM algorithm 2- Soft-classify users into LoS/NLoS 3- Reconstruct 3D map from “radio shadow” data Connections with -3D Imaging with WiFi [Y. Mostofi et al. @ UCSB] -GPS based 3D imaging [Madhow et al @ UCSB] [Esrafilian, Gesbert, Globecom 2017]
- 17. Joint 3D and radio map reconstruction 17 An orthoimagery of an area at center Washington DC, USA [Chen, Esrafilian, Gesbert, Mitra, Robotics, Science and Systems, MIT, 2017] Radio map reconstruction SNR/Radio map estimate 3D city map estimate Joint approach
- 18. DroneFor5GLab @ EURECOM Live video demo Please contact David G to watch it again! -