CorteXlab: research opportunities Jean-Marie Gorce jean-marie.gorce@insa-lyon.fr CorteXlab inauguration Tuesday, October 28, 2014
Which theoretical framework? EM theory Information theory Queuing theory Random matrices Game theory Stochastic Geometry 2
Information theory 1948 : Claude Shannon founded the Information Theory and provided, among many things his second theorem on system capacity: C Blog 1 SNR But this limit was achieved within 1bit in 1993 only!!! [turbo-codes, Berrou and Glavieux] 3
Information theory Information theory provides a strong theoretical framework Bits/J converse The capacity theorem relies on proving complementarily the converse and the achievability (from a theoretical point of view) Experimental research is nevertheless required to measure the gap between the upper-bound and an experimentally achievable capacity. Done Bits/Hz Asymptotically achievable 4
MIMO extension The theoretical capacity was virtually exceeded with MIMO systems. A long outstanding research extended the theoretical limit to these systems during the last 20 years (Foschini 1998, Telatar 1999, Gesbert 2002, ) But up to now, the real achievable capacity is not exactly known due to impairments, channel correlations, channel feedback,... Massive MIMO is a challenging research direction for 5G. 5
PHY layer perspectives Some open problems in point-to-point: Low complexity coding/decoding (LDPC, polar codes) Short codes [Poliansky, Poor, Verdu 2010] Channel or system impairments 6
PHY layer perspectives Two promising research directions are relative to: Multi-user communications: Resource sharing, interference alignment Cooperation (virtual MIMO, relay, spatial coding, ) Feedbacks Secrecy Software Defined Radio / Cognitive Radio More degrees of freedom Sensing, monitoring, sharing Cooperation Real-time and reconfigurability 7
Multi-user communications a general problem formulation
Multi-user communications Considering multi-user communications means to design, optimize or model simultaneous communications: Tx Rx Tx Rx Tx Rx PHY and MAC designed separately Cross-layer design: PHY and MAC 9
Network information theory (1) A Discrete Memoryless Network (DMN) is modeled as: [El Gamal & Kim 2012] X, k Y k X k Y, 1 k1 X, N Y N P ( y1 yn x1 xn ) X N Y, 1 N 1 1 1 X 2,Y 2 10
Network information theory (1) A Discrete Memoryless Network (DMN) is modeled as: [El Gamal & Kim 2012] ˆM 1 X, k Y k X k Y, 1 k1 X, N Y N P ( y1 yn x1 xn ) X N Y, 1 N 1 ˆM 1 1 1 M 1 X 2,Y 2 11
Network information theory (1) A Discrete Memoryless Network (DMN) is modeled as: [El Gamal & Kim 2012] ˆM 2 ˆM 1 X, k Y k X k Y, 1 k1 X, N Y N P ( y1 yn x1 xn ) X N Y, 1 N 1 ˆM 1 1 1 M 1 X 2,Y 2 M 1 12
Network information theory (1) ˆM 2 ˆM 1 X, k Y k A Discrete Memoryless Network (DMN) is modeled as: X k Y, 1 k1 P ( y1 yn x1 xn X, ) N Y N X N Y, 1 N 1 R js, D S j [El Gamal & Kim 2012] The total capacity is upper-bounded (cutset) by: c j 0 I X ( S); Y( S c ) X S C 1 1 X 2,Y 2 ˆM 1 A genie controls everything M 1 M 1 13
Network information theory (2) Even the simplest is hard! Which reference scenario? Multi-dimensional problem Capacity=polymatroid regions What is optimal? Fair or not fair? 14
Network information theory (3) Extensions OFDM/MIMO: // channels Non stationary channel Fire the genie? Secrecy: multi-nodes wire-tap channel Eve droppers are present Alice Tx Rx Bob Rx Eve 15
Network information theory (4) NIT provided numerous results for «small scale wireless networks» (SSWN) Several big steps forward, but still little is known: Relay channel (RC): D&F, C&F, Multiple Access channel (MAC,MARC) Broadcast channel (BC) Interference channel (IC) 16
What about Medium Scale WN? NIT associated with stochastic geometry may provide theoretical bounds for large scale wireless networks [Gupta&Kumar 2000], [Baccelli et al 2012]. But most practical systems are neither really small nor so large ~10 or 20 nodes medium what can we say about MSWN? Theory could be guided by experimental results 17
Perspectives with CorteXlab SSWN : CorteXlab may be used to test recent results in NIT : BC, MAC or IC scenarios Multiple nodes are easily and remotely accessible. The channel is controlled and almost Gaussian. Wired synch. and full feedback are available. The genie is here! MSWN: CorteXlab may be used to push forward original schemes Try and evaluate MU scenarios, relaying, cooperation, Define new retransmission/cooperation protocols. Fast exhaustive search on a stable Gaussian channel. 18
Distributed resource allocation (interference as noise) with V. Garcia, N. Lebedev, in assoc. with ALU
Distributed resource allocation A classical problem for : Ad hoc networks Cellular networks Different class of methods Control theory Game theory Ramdom fields (Gibbs sampling) User 1 User 2 20
The classical distributed power allocation problem [Foschini & Miljanic; 1993] [M. Chiang et al 2008] User 1 User 2 B base stations and 1 user/bs Power minimization: min i s. t. p i i e ; i i 21
Implementation in CorteXlab Features: Many links. TX1 TX2 TX3 Power control with feedbacks. External interference level. All layers (APP to PHY) Results: Convergence, stability, External Interference 22
Multi-resources distributed allocation Multiple Gaussian channels problem User 2 User 1 B base stations and 1 user/bs W : bandwidth divided in F frequency sub-bands Power minimization: min i s. t. r p r i r log 1 r C ; i i i Local water-filling Interference 23
Existing approaches The distributed water-filling problem is not convex. The objective is to reach a local minima. Solutions: Game theory based approaches: alternative selection of one user who optimizes its utility [Poor et al 2013] Gibbs sampling: requires partial information exchange between sources, trade between egoist and altruist objectives Control theory : small and controlled variations around the current solution (PhD Virgile Garcia, 2012) 24
P(r 2 ) Illustration with 2 resources Power trajectory Transmission in fading channels P(r 2 ) Interference trajectory P(r 1 ) P(r 1 ) P(r 2 ) Interference trajectory P(r 2 ) Power trajectory Successive water-filling Transmission in fading channels -Trade-off to build next step P(r 1 ) P(r 1 ) 25
Capacity (bps/hz) Interference (dbm) Resource 1 Resource 2 Obj. Mesure Prediction 26
Capacity (bps/hz) Interference (dbm) Resource 1 Resource 2 Obj. Mesure Prediction 27
Implementation in CorteXlab Features: OFDM based transmissions. TX1 TX2 TX3 Pilots, CSI. Synchronization Selective fading required. Dynamic channels. External Interference 28
Conclusion on dynamic and distributed decisions Dynamic and distributed decision in wireless networks is an important topic for increasing the global wireless capacity. Reference implementations are required to: - compare and evaluate different approaches. - take into account quantization, synchronization effects. - develop appropriate feedbacks and pilots. - prove stability and efficiency. 29
Interference alignment (interference as noise) with P. Ferrand, L. Cardoso In assoc. with Greentouch
Extension with Interference Alignment User 1 User 2 B base stations and U users U b : nb. of users attached to BS b W : bandwidth divided in F frequency sub-bands M antennas on the BSs, N on the UEs 31
NxF Principle of interference alignment MxF [Maddah,Motahan,Khandani, 2006] [Cadambe&Jaffar, 2008] H ii VFDM [Cardos et al 2013] G ij s ˆ1 D 1 H 11 C 1 s 1 D1 G1 b Cb sb D1 b1 N 32
Alternative model for cellular networks Pre-defined subspace [Suh & Tse 2011] K < NxF 33
Alternative model for cellular networks Pre-defined subspace [Suh & Tse 2011] K < NxF 34
Alternative model for cellular networks Pre-defined subspace [Suh & Tse 2011] K < NxF 35
Alternative model for cellular networks Pre-defined subspace [Suh & Tse 2011] K < NxF The BS selects the precoders s.t. C u C u W u' 36
Alternative model for cellular networks 37 [Suh & Tse 2011] Pre-defined subspace K < NxF N D s C G D s C H D s C H D s u b b U u u u ub u u u U u u u ub u u u ub u u b b ' ' ' ' ' ' ' ' ' ' ˆ
Implementation in CorteXlab Features: Each BS uses a subspace K over N OFDM/MIMO: a simplified OFDM frame Pilots, CSI, wired feedback. Synchronization BS2 BS1 Analysis: Robustness with real feedback loops Impact of quantization Convergence/equilibrium External Interference 38
Conclusion on interference alignment Interference alignment and similar techniques have a potential of a high throughput gain by exploiting phase diversity and creating interference holes. But they may suffer from high sensitivity to quantization, synchronization, estimation impairments. Reference implementations are required to evaluate robustness w.r.t. quantization/synch/estimation impairments 39
The 2-user-pair Symmetric IC (with feedback) with S. Perlaza, V. Quintero-Florez With R. Tandon, H.V. Poor
2-users-pair Symmetric IC (with feedback) [Ahlswede 1974] : IC model [Han & Kobayashi 1981] : tight inner bound [Tse & al 2011] : with feedback 41
Linear Deterministic IC (LD-IC) Four parameters (n 11,n 22,n 12,n 21 ) [Avestimehr et al 2007] with n ii log 2 SNR i nij log 2 INR ji 42
LD-IC without Feedback: Capacity Region [Bresler & Tse 2008] 43
Distributed decisions: Nash equilibrium What is the Nash Equilibrium Region of the Interference Channel with Feedback? 44
LD-IC without Feedback: Nash Region [Berry &Tse 2008] 45
LD-IC with perfect Feedback: Capacity & Nash Region [Suh & Tse 2011] [Perlaza, Tandon, Poor2012] 46
Example of achievability (5,5) at NE [Perlaza, Tandon, Poor2012] 47
Example of achievability (5,5) at NE [Perlaza, Tandon, Poor2012] 48
Example of achievability (5,5) at NE [Perlaza, Tandon, Poor2012] 49
Example of achievability (5,5) at NE [Perlaza, Tandon, Poor2012] 50
Example of achievability (5,5) at NE [Perlaza, Tandon, Poor2012] 51
Application in real situations? Validity without feedback: log log 2 2 I / N C / N log log 10 10 I / N IdB C / N CdB Interference as noise optimal if: I 0. 5 db C db Successive decoding optimal if: I 2 db C db TX1 TX2 Validity with feedback: A gain exists even in very weak or very strong regimes Cost of feedback? 52
Implementation in CorteXlab Features: AWGN channels Wired feedback available TX1 TX2 Outputs: Practical insights (coding, ) Find the different regimes Distributed decisions Cooperation, signaling 53
Implementation in CorteXlab Extension to K-users IC TX1 TX2 TX3 54
Conclusion on Interference channel Interference channel provides a potential capacity gain in a wide range of situations. An experimental evaluation would help to: - evaluate the real gain in a real scenario - design codebooks and cooperation mechanisms - design efficient distributed decision algorithms - study multiple nodes scenarios Same ideas could be addressed for BC, MAC and RC 55
Technological challenges Current and future works
Reference implementations Objective : provide the user with a pool of standardized functions (from PHY to application layers) A better exploitation of GNU radio community work Managing other environments s.a. Open Air Interface, Labview, Create a developers community dedicated to CorteXlab 57
Toward a realistic channel The current room is too perfect Good: Controlled, Static, no interference and ~LOS channel Bad: Static and lack of diversity Technical solution: Deployment of Static reflectors, attenuators and walls controllable mode stirrers JM.Gorce CorteXlab inauguration 58
Toward a realistic channel Software based channel control prefiltering + multiple sources S ( f ) C S 1 2 1 ( f ) C f X ( f ) f X ( f ) 2 prefilters may be used to control TX1 Create dynamic and controllable selective fading j2fdx/ c j2fdx'/ c h e C f h e C f X ( ) Y( f ) 1 2 2 1 f 59
Other perspectives inside CorteXlab Objective: Let the place for new equipments New RF front-ends may be installed for evaluation in a MU context. (why not mm-waves?) Deployment of robots is in preparation!!! Software radio node and FPGA remote access (under development) 60
Future projects We already have some projects on: K-nodes relay channels (with C&F schemes?) Distributed antenna systems and virtual MIMO Full duplex. 61
Future projects We expect collaborations or independent users of CorteXlab with various skills : RF and electronic design: test your own chip or RF in CorteXlab. Software radio: add new processing in FPGA or processor, design a very flexible node! Cognitive radio networks: manage them in a real environment. Signal processing: deploy your waveforms and/or coding. Signal processing: bring and share (or not) your own code on PC or FPGA. Information theory: validate new bounds or coding. 62
As a conclusion CorteXlab may foster your research by providing: A remotely accessible MU wireless testbed All software and hardware components you would not like to design yourself A controllable environment to challenge your designs with other teams. Any suggestion??? Help us to build the test bed you would like! 63
www.cortexlab.fr - 64