NetBeam: Network of Distributed Presenter: Carlos Bocanegra - - PowerPoint PPT Presentation

Presenter: Carlos Bocanegra

Next GEneration NEtworks and SYStems Lab

Advisor:

• Prof. Kaushik R. Chowdhury

NetBeam:

Network of Distributed Full-dimension Beamforming SDRs for Multi-user Heterogeneous Traffic

> <

2

Motivation

2. Boom of new data services - verticals 3. Novel deployments

Resource Management Entity Pedestrian

• riented

Urban blended UAV oriented sub-6GHz Antenna Array Aerial on urban oriented High mobility

• n the ground
• riented

1. + N. users, N. devices and data consumption 4. Wireless efficiency (IoT)

> <

3

Enablers

1. Beamforming or BF

! " ! ! !#$%(") (!#$%(") 3!#$%(")

> <

4

Enablers

1. Beamforming or BF 2. Distributed and Collaborative Beamforming (DCBF)

> <

5

Enablers

1. Beamforming or BF 3. Distributed Antenna Systems (DAS) 2. Distributed and Collaborative Beamforming (DCBF)

> <

6

Enablers

4. Full dimensional BF or 3DBF 1. Beamforming or BF 3. Distributed Antenna Systems (DAS) 2. Distributed and Collaborative Beamforming (DCBF)

*AAS: Active Antenna Aystem

> <

7

Scenario

DAS deployment: antennas are coordinated and synchronized via RU. Heterogeneous Users: 3D locations and demands (SINR/QoS). Multi-User comms: Multiple users served using NOMA. Electronic tilt: Each antenna modifies its azimuth/elevation angles.

Resource Management Entity

z x y ! ∅ 3-D antenna steering

Distributed Antenna Array

Group 1 Group 2 Group 3 3-D distributed users Disimilar traffic demands

𝒊𝒋 Channel user i 𝜄 Antenna elevation 𝜒 Antenna azimuth 𝒙𝒋 Beamforming weights user i 𝚳 𝒋 QoS for user I (SINR) 𝑸𝒖 𝒐 Maximum Transmit Power

PROBLEM 1: What are the angles 𝜄, 𝜒 that maximize the channel gain per each antenna pair? PROBLEM 2: (1) solution is NP-Hard. Need constraint relaxation.

System Goal Formulation

Total power Achieved SINR for user i Requested SINR for user i Restriction on available power

𝑅𝑝𝑇H 𝑅𝑝𝑇I 𝑅𝑝𝑇J

How to select azimuth/elevation angles, group antennas and beamforming weights to minimize the transmit power while ensuring demanded QoS per user.

> <

• Helps reducing the power consumption at the transmitter and rises the channel efficiency.
• Antenna selection under dissimilar application demands (QoS/SINR).

8

Related Work

[10] N.Seifi et al., “Adaptive Multicell 3-D Beamforming in Multiantenna Cellular Networks,” Tech. Rep. 8, 2016. [12] N. M. Boers and D. Mak, “Impact of orientation and wire placement on received signal strengths,” in IEEE CAMAD, 2016.

• 3DBF to increase the sum rate via channel decorrelation or higher channel gains.
• The omnidirectionality of antennas is not perfect and can be exploited.

[17] G. Sun et al., “A Sidelobe and Energy Optimization Array Node Selection Algorithm for Collaborative Beamforming in Wireless Sensor Networks,” IEEE Access, 2017. [18] B. Bejar Haro et al., “Energy efficient collaborative beamforming in wireless sensor networks,” IEEE Transactions on Signal Processing (TSP), 2014.

• Node placement may not be available in some scenarios, i.e. restricted areas.
• Do not encompass dissimilar traffic demands (applications).

[8] Y. Gao et al., “Massive MIMO Antenna Selection: Switching Architectures, Capacity Bounds, and Optimal Antenna Selection Algorithms,” IEEE TSP, 2018. [15] Kyungchul Kim et al., “Spatial-Correlation- Based Antenna Grouping for MIMO Systems,” IEEE Transactions on Vehicular Technology, 2010.

1. Full dimensional beamforming and antenna steering 2. Beamforming in DCBF/ DAS systems 3. Antenna selection

> <

9

Proposed framework - NetBeam

Distributed Antennas Wireless connection Heterogeneous users Antenna selection Socket programming Feedback and control Data, control, and signaling Transmitter settings and control Resource Management Entity 3DBF Time Sync. MCS selection Antenna steering Time Sync. SDR Beamforming TX Beamforming RX Time Sync. Time Correction Channel Estimation Feedback interface Time correction, Channel Estimation Beamforming Connection awareness Distributed Base Stations

> <

10

Proposed algorithm - NetBeam

𝑵∗

Λ

𝒊𝟐 EGO-based Antenna Orientation Spatial interpolation: Kriging Prior: - Gaussian fit Trial selection: DIRECT-UM 𝒊𝟑 𝒊𝟒 𝒊𝑶

𝜄∗, 𝜒∗

Antenna Selection via Matching Modified Hungarian alg. Binary search space

𝜄QRH, 𝜒QRH

𝑰T 𝑰∗

Efficient Digital Beamforming Semi-Positive Definite Power minimization at Tx.

𝑿∗ Ω

User’s app. Demands: SNR 𝒊𝒋 Channel user i

𝜄

Antenna elevation

𝜒

Antenna azimuth

𝑰T Maximum gain channel

𝑵∗

Assignation matrix

𝑰∗

Optimum channel

Ω

Angular space

Λ

SNR demands

𝑿∗ Optimum digital weights

NetBeam Overview

• Sec. V
• Sec. VI
• Sec. VII

*EGO: Efficient Global Optimization

> <

11

Antenna orientation (AO)

Distributed Antennas Wireless connection Heterogeneous users Antenna selection Socket programming Feedback and control Data, control, and signaling Transmitter settings and control Resource Management Entity 3DBF Time Sync. MCS selection Antenna steering Time Sync. SDR Beamforming TX Beamforming RX Time Sync. Time Correction Channel Estimation Feedback interface Time correction, Channel Estimation Beamforming Connection awareness Distributed Base Stations

> <

12

AO – Preliminary studies

15 30 45 60 75 90 30 60 90 120 150 180

Find the angles that return the best RSSI for each user in the minimum number of trials.

Goal:

• A. Model the angular map.
• B. Select next angular tuple to evaluate.
• C. Reached max trials? -> Return tuple
• D. Else -> return to A.

*Angular map: Azymuth and Elevation angles at the X-Y plane, RSSI at the Z-plane

Approach: Sequential decisions problem

15 30 45 60 75 90 30 60 90 120 150 180 15 30 45 60 75 90 30 60 90 120 150 180 15 30 45 60 75 90 30 60 90 120 150 180 15 30 45 60 75 90 30 60 90 120 150 180

Efficient Global Optimization framework (EGO)

Gaussian Processes and Kriging

Novel DIRECT-UM

𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐵𝑨𝑗𝑛𝑣𝑢ℎ

> <

13

Modeling the angular map

AO – Gaussian Processes

𝑢j = 𝑧j + 𝜗j = 𝑔(𝑦j) + 𝜗j 𝑧j 𝑦j

channel AWGN

𝑢j 𝑔(𝑦j): The wireless channel 𝜗j: Additive White Gaussian Noise (AWGN) 𝜗j~𝑂(0, 𝛾wH) 𝑄(𝑢j|𝑧j)~𝑂(𝑢j, 𝛾wH z 𝜀) 𝑄(𝑧j|𝑦j)~𝑂(𝑧j, 0 z 𝑳) AWGN characterization: Channel characterization:

Find a precise estimate of 𝒖𝒐 keeping n low 𝒖𝒐 → {𝒚𝟏, 𝒚𝟐, 𝒚𝟑,…, 𝒚𝒐w𝟐}

𝑧j 𝑢j 𝜸w𝟐 𝒖𝒐 𝑄(𝑢j)~ ‚ 𝑄 𝑢 𝑧 𝑄 𝑧 𝑒𝑧 𝑳 𝑦j, 𝑦„ : Kernel or covariance Channel characterization: 𝑂(𝑧j, 0 z 𝑳) 𝑂(𝑢j, 𝛾wH z 𝜀) 𝑫 𝑦j, 𝑦jwH = 𝑳 𝑦j, 𝑦jwH + 𝛾wH z 𝜀j„ 𝑄(𝑧|𝑧†)~𝑂 𝜈 𝜈† , 𝐷 𝐷∗ 𝐷∗

‰

𝐷† Posterior distribution (1 observation):

15 30 45 60 75 90 30 60 90 120 150 180

𝑦j

𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜

> <

14

Modeling the angular map

AO – Gaussian Processes regression (Kriging)

𝑢j = 𝑧j + 𝜗j = 𝑔(𝑦j) + 𝜗j 𝑧j 𝑦j

channel AWGN

𝑢j 𝑔(𝑦j): The wireless channel 𝜗j: Additive White Gaussian Noise (AWGN)

Find a precise estimate of 𝒛𝒐 keeping n low 𝒛𝒐 → {𝒚𝟏, 𝒚𝟐, 𝒚𝟑,…, 𝒚𝒐w𝟐}

15 30 45 60 75 90 30 60 90 120 150 180

𝑦j

Multi-observation: Gaussian Process Regression (GPR) 𝒛𝒐 | {𝒚𝟏, 𝒚𝟐, 𝒚𝟑,…, 𝒚𝒐w𝟐} ?

𝑧j = 𝑥H𝑦H + 𝑥H𝑦H + ⋯ + 𝑥H𝑦H + 𝑕 + 𝜁j = 𝐗𝐔𝐲 + 𝐡 + 𝜻

20 40 60 80 100 120 140 160 180

Azimuth antenna steering

0.008 0.01 0.012 0.014 0.016 0.018 0.02 0.022

Empirical channel gain

Kriging-based prediction using 4 trials

Real measurement Trials (known) Prediction (mean) Uncertainty (variance)

𝑒H 𝑒J 𝑒” 𝑒I 𝑋

H > 𝑋 J > 𝑋 ”> 𝑋 I

𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜

Kernel selection – The exponential semivariogram

𝐿 ℎ 𝑑†, 𝑏† ~ 𝑑† 1 − exp(− ℎI 𝑏†

I)

𝑢›?

> <

15

• DIRECT starts with an off-line stage, exploring all the points in the set.
• DIRECT aims to expand symmetrically on the desired region.

AO – DIRECT for trial selection

Goal: trial selection

15 30 45 60 75 90 30 60 90 120 150 180 15 30 45 60 75 90 30 60 90 120 150 180 15 30 45 60 75 90 30 60 90 120 150 180 15 30 45 60 75 90 30 60 90 120 150 180

Divide in RECTangles (DIRECT)

15 30 45 60 75 90

Elevation

30 60 90 120 150 180

Azimuth 6 trials

𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜

> <

15 30 45 60 75 90 30 60 90 120 150 180

#6 kriging variance

16

AO – DIRECT-UM for trial selection

15 30 45 60 75 90 30 60 90 120 150 180

#8 kriging variance

15 30 45 60 75 90 30 60 90 120 150 180

#5 kriging variance

15 30 45 60 75 90 30 60 90 120 150 180

#4 kriging variance

DIRECT-UM = DIRECT + Uncertainty minimization (UM)

𝑶 = 𝟓 𝑶 = 𝟔 𝑶 = 𝟕 𝑶 = 𝟗

𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜

𝑷𝑮𝑮 − 𝑴𝑱𝑶𝑭 𝑷𝑶 − 𝑴𝑱𝑶𝑭

> <

17

Antenna selection

Distributed Antennas Wireless connection Heterogeneous users Antenna selection Socket programming Feedback and control Data, control, and signaling Transmitter settings and control Resource Management Entity 3DBF Time Sync. MCS selection Antenna steering Time Sync. SDR Beamforming TX Beamforming RX Time Sync. Time Correction Channel Estimation Feedback interface Time correction, Channel Estimation Beamforming Connection awareness Distributed Base Stations

> <

18

Antenna selection

𝑢H 𝑢I 𝑢J 𝑢” 𝑢› 𝑠

H

𝑠

I

𝑠

J

𝐷HH 𝐷›J

Assign Antennas from set T to users in set R to maximize the sum Rate

Goal:

Hungarian algorithm – Polynomial time in 1-to-1 assignations form T to R Perfect Matching problem Modi§ied Hungarian: 𝑃(2 ∗ 𝐿 ∗ 𝑈 J) Hungarian: 𝑃( 𝑈 J) Modified Hungarian algorithm is still Polynomial

𝑠

H

𝑠

H ¬

𝑠

H ¬¬

𝑠

H ¬¬¬

𝑢H 𝑢I 𝑢› 𝑠

I

𝑠

I ¬

𝑠

I ¬¬

𝑠

I ¬¬¬

𝑠 𝑠

J ¬

𝑠

J ¬¬

𝑠

J ¬¬¬

…

> <

19

Low-complexity BF weights

Distributed Antennas Wireless connection Heterogeneous users Antenna selection Socket programming Feedback and control Data, control, and signaling Transmitter settings and control Resource Management Entity 3DBF Time Sync. MCS selection Antenna steering Time Sync. SDR Beamforming TX Beamforming RX Time Sync. Time Correction Channel Estimation Feedback interface Time correction, Channel Estimation Beamforming Connection awareness Distributed Base Stations

> <

20

Low-complexity BF weights

• The problem in (1) is non-convex and requires additional modifications to

find a global optima un the system:

Use of the channel correlation matrix R

The inclusion of the channel correlation matrix R and the removal of the constraint over the unitary rank for W still doesn’t lead to false optimum.

• Semi-Positive Definite (SDP) relaxation allows us to

compute (1) in Polynomial time via constraint relaxation

> <

21

Device synchronization

Distributed Antennas Wireless connection Heterogeneous users Antenna selection Socket programming Feedback and control Data, control, and signaling Transmitter settings and control Resource Management Entity 3DBF Time Sync. MCS selection Antenna steering Time Sync. SDR Beamforming TX Beamforming RX Time Sync. Time Correction Channel Estimation Feedback interface Time correction, Channel Estimation Beamforming Connection awareness Distributed Base Stations

> <

TX1 TX2 RX1

• Rx detects indices after correlation and establish mismatch between Rx GS.
• Rx reports number of samples one of the transmitters should pre-pend its signal with CSI.

(T echnique usually called Time Advanced or TA)

Correction reported using the out-of-band

WiFi Access Point

802.11ac

Frequency sync. Phase and Time sync. 1st Tx should defer its transmission Synchronized

22

Device synchronization and Frame structure

3. Frame synchronization

> <

23

Experimental setup

Servo Motor 1

TX1 TX2

Tx Radio

RX1

Servo Motor 2

Tx Radio Rx Radio

LoS Configuration angle

• X310 radios as TX and B210 as RX, operating in the 900MHz band.
• Each antenna is assigned a Gold Sequence (Good cross-correlation) for synchronization and CSI purposes.
• The CSI and correlation indexes are shared using the out-of-band 2.4GHz link over the TP-link router.

> <

• The graphs show the channel gain for different Elevation and Azimuth pairs employing an exhaustive search.
• The outdoor environment has the LoS as its optimum orientation, mainly due the absence of scatterers.
• The indoor environment reveals a bigger difference between LoS and optimum channel gain.

15 30 45 60 75 90 30 60 90 120 150 180 15 30 45 60 75 90 30 60 90 120 150 180 15 30 45 60 75 90 30 60 90 120 150 180 15 30 45 60 75 90 30 60 90 120 150 180 15 30 45 60 75 90 30 60 90 120 150 180 15 30 45 60 75 90 30 60 90 120 150 180

• Max. gain

LoS

15 30 45 60 75 90 30 60 90 120 150 180 15 30 45 60 75 90 30 60 90 120 150 180 15 30 45 60 75 90 30 60 90 120 150 180 15 30 45 60 75 90 30 60 90 120 150 180 15 30 45 60 75 90 30 60 90 120 150 180 15 30 45 60 75 90 30 60 90 120 150 180

• Max. gain

LoS

Indoors Outdoors

24

AO – Preliminary studies

𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜

> <

25

We apply a Gaussian fit to

• ur experimental data

along with the Best Linear Unbiased Estimator (BLUE), which offers the lowest variance. The spatial correlation is done vie the semi- variogram.

50 100 150

Angular distance

0.01 0.02 0.03 0.04 0.05 0.06

(h)

50 100 150

Angular distance

0.02 0.04 0.06 0.08 0.1

(h)

Rx 2 Tx 5: Empirical Variogram Rx 2 Tx 6: Empirical Variogram Rx 2 Tx 7: Empirical Variogram Rx 2 Tx 8: Empirical Variogram Rx 2 Tx 5: Analytical Rx 2 Tx 6: Analytical Rx 2 Tx 7: Analytical Rx 2 Tx 8: Analytical

< 54° , 140° > Window < 75° , 98° > < 64° , 98° > < 73° , 113° > height: 45'' height: 73'' height: 51'' height: 73'' d: 40'' d: 40'' d: 60'' d: 10''

Receiver Transmitter

d: 52'' height: 29'' height: 29'' height: 29'' RX1 RX2 RX3 TX5 TX7 TX8 TX6 Open area < 64° , 98° > height: 51'' < 75° , 98° > < 73° , 113° > < 54° , 140° > height: 45'' height: 73'' height: 73'' d: 40'' d: 40'' d: 60'' d: 10''

Receiver Transmitter

d: 52'' Concrete wall height: 29'' height: 29'' height: 29'' RX1 RX2 RX3 TX5 TX7 TX8 TX6

Configuring our Kernel

Indoor Locations Outdoor Locations An angular distance ||𝑀||I of 100 is a taken as the limit for meaningful correlation. OUTDOORS INDOORS

> <

26

Performance evaluation

The proposed DIRECT-UM closes the gap to optimality within a reduced number of trials. The On-line stage selects trials intelligently depending on the uncertainty.

Random DIRECT-RD DIRECT-UM PI UM 0.02 0.04 0.06 0.08 0.1

Gap to optimality (linear gain) OUTDOORS

Off-line: 2 trials / On-line: 8 trials Off-line: 4 trials / On-line: 8 trials Off-line: 6 trials / On-line: 8 trials Off-line: 8 trials

On-line stage in DIRECT-UM minimizes the Gap to

• max. channel gain

Random DIRECT-RD DIRECT-UM PI UM 0.05 0.1 0.15 0.2

Gap to optimality (linear gain) INDOORS

Off-line: 2 trials / On-line: 8 trials Off-line: 4 trials / On-line: 8 trials Off-line: 6 trials / On-line: 8 trials Off-line: 8 trials

On-line stage in DIRECT-UM minimizes the Gap to

• max. channel gain

DIRECT-UM vs other acquisition functions 2. Uncertainty Minimization (UM) 1. Probability of Improvement (PI) 5. DIRECT with UM selection (Proposed) 3. Random 4. DIRECT with Random selection

15 30 45 60 75 90 30 60 90 120 150 180

𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜

*Gap to optimality: Gain difference between the known

• ptimum 𝑔(𝑦(𝜒∗, 𝜄∗)) and the

current chosen maximum 𝑔(𝑦(𝜒„, 𝜄„))

> <

27

Performance evaluation

• ptimum

greedy random

• 0.15
• 0.1
• 0.05

Overall SNR lost (linear) OUTDOORS

DIRECT-UM UM PI DIRECT DIRECT-RD Random

• ptimum

greedy random 2 4 6 8 10

Overall tx. power (linear) OUTDOORS

DIRECT-UM UM PI DIRECT DIRECT-RD Random

DIRECT-UM with Optimum antenna allocation

NetBeam as the optimum configuration 1. Modified Hungarian algorithm (Optimum) 2. Greedy selection 3. Random selection 2. Uncertainty Minimization (UM) 1. Probability of Improvement (PI) 5. DIRECT with UM selection (Proposed) 3. Random 4. DIRECT with Random selection Antenna steering policies Antenna selection policies NetBeam (multi-stage) proves to minimize the transmit power while still meeting the SINR requested per user.

> <

28

Conclusions

1. Designed a multi-stage optimization framework for Distributed Array Systems (DAS) serving users with heterogeneous traffic demands (SNR/SINR) and minimizing the total transmit power advocating for green environments in congested deployments. 2. (1st stage) Presented a ML-based antenna orientation mechanism exploiting 3DBF to find the max. channel gain in a per-user basis. 3. (2nd stage) Devised a transmitter antenna-to-user matching algorithm using perfect matching algorithms based on the CSI. 4. (3rd stage) Computed the digital beamforming weights using Semi-Definite Positive (SDP) techniques, aiming to min. the transmit power while serving each user with the requested QoS. 5. Validated the framework via implementation of a fully programmable distributed beamforming testbed using SDR’s operating in the ISM frequency band, with controlled out-of-band CSI feedback.

Thanks

APPENDIX

> < Servo Motor 1

TX1 TX2

Tx Radio

RX1

Goal: Does steering increase channel gain?

• Antennas are connected to servo-motors.
• Antennas are given orthogonal codes to

compute the RSSI at RX simultaneously.

• We average the RSSI over the multiple packets

sent per antenna angle pairs.

Servo Motor 2

Tx Radio Rx Radio Elevation: 90 degrees. Azimuth: 30 degrees. Elevation: 90 degrees. Azimuth: 150 degrees.

LoS Configuration angle

31

AO – Preliminary studies

> <

32

Gaussian regression for channel gain in angular domain

AO – Gaussian Processes regression (Kriging)

20 40 60 80 100 120 140 160 180

Azimuth antenna steering

0.008 0.01 0.012 0.014 0.016 0.018 0.02 0.022

Empirical channel gain

Kriging-based prediction using 4 trials

Real measurement Trials (known) Prediction (mean) Uncertainty (variance)

20 40 60 80 100 120 140 160 180

Azimuth antenna steering

0.008 0.01 0.012 0.014 0.016 0.018 0.02 0.022

Empirical channel gain

Kriging-based prediction using 6 trials

Real measurement Trials (known) Prediction (mean) Uncertainty (variance)

20 40 60 80 100 120 140 160 180

Azimuth antenna steering

0.01 0.012 0.014 0.016 0.018 0.02

Empirical channel gain

Kriging-based prediction using 13 trials

Real measurement Trials (known) Prediction (mean) Uncertainty (variance)

20 40 60 80 100 120 140 160 180

Azimuth antenna steering

0.01 0.012 0.014 0.016 0.018 0.02

Empirical channel gain

Kriging-based prediction using 7 trials

Real measurement Trials (known) Prediction (mean) Uncertainty (variance)

Then, how to select the next trial?

𝑶 = 𝟓 𝑶 = 𝟕 𝑶 = 𝟐𝟒 𝑶 = 𝟖

> <

33

AO – Preliminary studies

15 30 45 60 75 90 30 60 90 120 150 180

Find the angles that return the best RSSI for each user in the minimum number of trials.

Goal:

*Angular map: Azymuth and Elevation angles at the X-Y plane, RSSI at the Z-plane

15 30 45 60 75 90 30 60 90 120 150 180 15 30 45 60 75 90 30 60 90 120 150 180 15 30 45 60 75 90 30 60 90 120 150 180 15 30 45 60 75 90 30 60 90 120 150 180

𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜 𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐵𝑨𝑗𝑛𝑣𝑢ℎ

• A. Model the angular map.
• B. Select next angular tuple to evaluate.
• C. Reached max trials? -> Return tuple
• D. Else -> return to A.

Approach: Sequential decisions problem

Efficient Global Optimization framework (EGO)

Gaussian Processes and Kriging

Novel DIRECT-UM

20 40 60 80 100 120 140 160 180

Azimuth antenna steering

0.008 0.01 0.012 0.014 0.016 0.018 0.02 0.022

Empirical channel gain

Kriging-based prediction using 4 trials

Real measurement Trials (known) Prediction (mean) Uncertainty (variance)

15 30 45 60 75 90 30 60 90 120 150 180

#6 kriging variance

𝐵𝑨𝑗𝑛𝑣𝑢ℎ 𝐹𝑚𝑓𝑤𝑏𝑢𝑗𝑝𝑜

> <

I F F T

1 2 3 4 5 6 7 (N-1)/2 … (N-1)/2

+ + + + +

W

G.S. #2 G.S. #1 G.S. #4 G.S. #3 Generate Bits 64-QAM 32-QAM 16-QAM 8-QAM QPSK BPSK

900 MHz

Zero Padding +

34

Device synchronization and Frame structure

1. Frame structure – the Transmitter

…

XCORR G.S. #1 Synchronization + CSI + TA GS Set F F T

1 2 3 4 5 6 7

• (N-1)/2

… (N-1)/2

H Time Correction

≅ 𝟏

Known Bits Bit Error Rate

BER …

64-QAM 32-QAM 16-QAM 8-QAM QPSK BPSK XCORR G.S. #1 XCORR G.S. #1 XCORR G.S. #1

2. Frame structure – the Receiver