In recent years there has been a rapid increase in the number of wireless devices for both commercial and defense applications. Such unprecedented demand has increased device cost and complexity and also added a strain on the spectrum utilization of wireless communication systems. This thesis addresses these issues, from an antenna system perspective, by developing new techniques to dynamically optimize adaptive beamforming arrays for improved anti-jamming and reliability. Available frequency spectrum is a scarce resource, and therefore increased interference will occur as the wireless spectrum saturates. To mitigate unintentional interference, or intentional interference from a jamming source, antenna arrays are used to focus electromagnetic energy on a signal of interest while simultaneously minimizing radio frequency energy in directions of interfering signals. The reliability of such arrays, especially in commercial satellite and defense applications, can be addressed by hardware redundancy, but at the expense of increased volume, mass as well as component and design cost. This thesis proposes the development of new models and optimization algorithms to dynamically adapt beamforming arrays to mitigate interference and increase hardware reliability. The contributions of this research are as follows. First, analytical models are developed and experimental results show that small antenna arrays can thwart interference using dynamically applied stochastic algorithms. This type of in-situ optimization, with an algorithm dynamically optimizing a beamformer to thwart interference sources with unknown positions, inside of a anechoic chamber has not been done before to our knowledge. Second, it is shown that these algorithms can recover from hardware failures and localized faults in the array. Experiments were performed with a proof-of-concept four-antenna array. This is the first hardware demonstration showing an antenna array with live hardware fault recovery that is adapted by stochastic algorithms in an anechoic chamber. We also compare multiple stochastic algorithms in performing both anti-jamming and hardware fault recovery. Third, we show that stochastic algorithms can be used to continuously track and mitigate interfering signals that continuously move in an additive white Gaussian noise wireless channel.

1. Dynamic Beamforming Optimization for
Anti-Jamming and Hardware Fault Recovery
Jonathan Becker
Ph.D. Candidate, Electrical and Computer Engineering
Carnegie Mellon University
Thesis Advisor: Prof. Jason Lohn
Thesis Committee: Prof. Ole Mengshoel, Prof. Patrick Tague,
Dr. Derek Linden (CTO, X5 Systems, Inc.)

2. About Me
Jonathan Becker
15 years of research & industry experience in
machine learning, stochastic optimization,
antenna design, RFID wireless sensing, and
RF / microwave engineering design.
8 papers in the related fields.
Carnegie Mellon University / Ph.D. (2009-2014)
Advisor: Prof. Jason Lohn
University of Southern California / MSEE 2004
Cal Poly San Luis Obispo / BSEE with CS Minor 1999
Work Experience
Disney / Wireless Displacement Sensing (2012-2013)
EDO / Interference Cancellation Systems (2001-2006)
Teradyne / High-bandwidth IC tester interfaces (1999-2001)
2

3. Main Goal
The main goal of this research is to develop
foundational models of and stochastic
algorithms for anti-jamming beamforming in the
presence of static and mobile signals and
hardware faults.
3

4. Dynamic Beamforming Optimization With Fault Recovery: Motivation
Jammer
Desired
Jammer
Wireless Comm. Blocked
Array
Failure
BFFault
X
X
Anti-Jamming Beamforming
4
HW Fault
Recovery
via Alg.BFFault
No HW
Redundancy
Limited
Spectrum
BF
Volume
Constrained
X
X

5. Fault Tolerance Importance in Anti-Jamming Beamforming
5[1] H.H. Khatib, “Theater wideband communications,” IEEE MILCOM 97
Proceedings, pp. 378-382, 2-5 Nov. 1997.
• Failure to anti-jam can cause a ripple effect down the communication path.
• Reconfiguration of array weights during recovery provides anti-jamming
beamforming by definition Array Failure

6. Outline
• Motivation
• Previous Research
• Research Problems and Solutions
• Research Approach
• Experiments and Results
• Conclusion
6

7. Previous Research
7[1] D. Linden. “Optimizing signal strength in-situ using an
evolvable antenna system,” NASA/DOD EH Conf., 2002.
Reed Relay
Switch
Feed Point
RF Traces
GA optimized mainbeam gain by turning switches on and off

8. Previous Beamforming Research
• Haupt used 128 antennas to null one jammer
• Array tuned to signal of interest prior
8
Haupt’s antenna array One bank of attenuators
and phase shifters
7 Degrees of Freedom.
[1] R. Haupt and H. Southall, “Experimental adaptive nulling with a genetic
algorithm,” Microwave Journal, vol. 42, no. 1, pp. 78–89, 1999.

9. Previous Fault Tolerance Research
9
[1] Lee et. al., “A built-in performance-monitoring/fault isolation and
correction (PM/FIC) system for active phased arrays, IEEE
Transactions on Antennas and Propagation, Nov. 1993.
• 8 X 10 antenna active array for radar (mainbeam scanning)
• Injection of external signal for fault detection
• Complex circuitry needed to detect faults & re-tune array
Transmission
Line Injection
Control
Circuitry

10. Previous Fault Tolerance Research
• Han et. al. showed GA’s ability to resynthesize
beam pattern after transmit/receive module failed
10
J. H. Han, S. H. Lim, and N. H. Myung, “Array antenna TRM failure
compensation using adaptively weighted bean pattern mask based on
genetic algorithm,” IEEE Antennas and Wireless Propagation Letters, 2012.
GA reconfigured weights using
pattern mask based fitness function
AF = Array Factor

11. Previous Fault Detection Research
• Oliveri et. al. developed a fault detection approach
based on Bayesian Compressive Sensing
11
Oliveri et. al., “Reliable Diagnosis of Large Linear Arrays – A Bayesian
Compressive Sensing Approach,” IEEE Transactions on Antennas and
Propagation, October 2012.
ˆf = argmax
f
P f F( )!
"
#
$
“Difference”
field pattern P f F( )=
P F f( )P f( )
P F( )
Solve using
Bayes Theorem
Sparse
“failure” vector

12. Outline
• Motivation
• Previous Research
• Research Problems and Solutions
• Research Approach
• Experiments and Results
• Conclusion
12

13. Research Problems in Anti-Jamming Beamforming
Research problems:
1. Signal directions often time-varying
and unknown a priori, so canonical
beamforming techniques used with
Radar arrays not applicable.
2. How to search a large, combinatorial
parameter space with multimodal
fitness landscape?
3. How well do stochastic algorithms
adapt to mobile signals?
• Available frequency spectrum is a scarce resource.
• Increased interference will occur as the wireless spectrum saturates.
• Antenna arrays used to focus electromagnetic energy on a desired
signal of interest & minimize energy towards interfering signals
Goal: Perform anti-jamming in presence of static & mobile signals
13
BF
X
X
Anti-Jamming Beamforming

14. Fault Recovery Research Problems
Research problems:
1. Recovery from hardware failures
and localized faults in the array
2. How do stochastic algorithms treat
hardware faults vs. mobile
signals?
3. What happens if a hardware
component fails before algorithmic
convergence? After convergence?
• Hardware redundancy addresses antenna array reliability at expense of
more volume, mass, cost.
• Volume, mass, and cost constraints create lack of hardware redundancy.
• Faulted hardware components cause loss of anti-jamming functionality.
Goal: Perform HW fault recovery with stochastic algorithms
Fault
Recovery
BF
X
X
Fault
14

15. Anti-Jamming Beamforming Arrays
15
Shape radiation pattern using
multiple antennas and hardware
amplitude / phase weights
Null shifted to 30°

16. Stochastic Search Algorithms
16
Approach Features Drawbacks
Least Mean
Squares
Adaptive feedback
Local search with poor
multi-modal performance
Conjugate
Gradient
Method
Searches parameter
space using
conjugate directions
Signal directions needed,
poor multi-modal
performance, O(N2)
Genetic
Algorithms
Population based
global search
Run-time is problem
dependent
Simulated
Annealing
Evaluates solutions
sequentially
Convergence is cooling
schedule dependent

17. Outline
• Motivation
• Previous Research
• Research Problems and Solutions
• Research Approach
• Experiments and Results
• Conclusion
17

18. Triallelic Diploid Genetic Algorithm
18
SINROut tn,p
!" #$=
PS,Out tn,p
!" #$
Pj,Out tn,p
!" #$+ No
j=1
J
∑
Fitness Function:
SINR = Signal to Interference and Noise Ratio

19. Simple Genetic Algorithm
19
SINROut tn,p
!" #$=
PS,Out tn,p
!" #$
Pj,Out tn,p
!" #$+ No
j=1
J
∑
Fitness Function:
SINR = Signal to Interference and Noise Ratio

20. Simulated Annealing Block Diagram
20
SINROut tn[ ]=
PS,Out tn[ ]
Pj,Out tn[ ]+ No
j=1
J
∑
Fitness Function:

21. Hill Climbing Block Diagram
21
SINROut tn[ ]=
PS,Out tn[ ]
Pj,Out tn[ ]+ No
j=1
J
∑
Fitness Function:

22. Wireless Channel Model
22
Symbol Meaning
J Number
jammers
Q Number
reflections
N Number
antennas
• Array creates single weighted
sum of signals and reflections
• Signal directions unknown a priori
• Sum of signals and multipath
reflections should not exceed
number of antennas in the array
• Array response calculation
important for simulation fidelity
N antennas
Antenna Array Response

23. Array Factor Model of Antenna Arrays
P(R,θ,ϕ)
• Models antennas as infinitesimal dipoles
• Far-field computation in O(N) time
• Ignores antenna mutual coupling and reflections off objects near antennas
23
AF θ,φ( )= ˆaiejβaR⋅di
i=1
N
∑
ˆai = aie− jψi
ai ∈ ℜ and 0 < ai ≤1
ψi ∈ ℜ and 0 ≤ ψ < 2π
Complex array
weights
Element Positions
Spherical Unit Vector

24. Method of Moments (MOM) Model of Antenna Arrays
• Models antennas as combination of small wire segments
• Mutual coupling included in calculation of far-field radiation patterns
• Ignores reflections off objects near antennas
24
Fields calculated in O(N3) time.
Goal: Given known port excitations, solve
integral equations to calculate
currents on each wire
Solution: Divide each wire into segments and
estimate unknown currents as sum
of weighted basis functions
Segmentation
on N antennas
ˆI = ˆZ!
"
#
$
−1
ˆV
Result: Ultimately obtain vector equations of
form
Post-processing: Calculate far-fields
D. B. Davidson, Computational Electromagnetics for RF and Microwave
Engineering, 2nd ed. New York, NY: Cambridge University Press, 2011
Matrix
Inversion

25. Antenna Arrays with Nearby Objects
• Physical arrays include metallic objects near antennas
• Incorporate reflections into MOM by including metallic objects in model
• Model objects as Perfect Electric Conducting (PEC) planes
25Fields calculated in O(N3) time.

26. Optimal Array Weights with Mutual Coupling
Compensation
26
MC
−1 ˆaM,opt = MC
−1 ˆaAF,opt
ˆaAF,opt
Inverse of
Coupling Matrix
MC found using MOM compared to array factor calculation
Optimized weights
using Array Factor
calculations
Coupling compensated
optimized weights
[1] T. Zhang and W. Ser, “Robust beampattern synthesis for antenna arrays with mutual coupling effect,”
IEEE Transactions on Antennas and Propagation, vol. 59, no. 8, pp. 2889–2895, 2011
[2] P. J. Bevelacqua, “Antenna arrays: Performance limits and geometry optimization,” Ph.D. dissertation,
Arizona State University, May 2008.
[3] M. Joler, “Self-recoverable antenna arrays,” IET Microwaves Antennas Propagation, vol. 6, no. 14, pp.
1608–1615, 2012.

27. Optimal Array Weights with Mutual Coupling
and Hardware Reflection Compensation
27
MCR
−1 ˆaMR,opt = MCR
−1 ˆaAF,opt
ˆaAF,opt
Inverse of
Coupling Matrix
New Method: Hardware reflections
compensation not discussed in literature
Optimized weights
using Array Factor
Coupling + Reflection
compensated optimized
weights
MR
−1 ˆaMR,opt = MR
−1 ˆaM,opt
ˆaM,opt
Inverse of
Reflection Matrix
Coupling + Reflection
compensated optimized
weights
MCR = MC MR
MOM
output

28. Equivalence of Stochastic Algorithms with
Different Antenna Array Models
28
Need to calculate an inverse matrix for each transformation
Most Reliable but O(N3)Least Reliable but O(N)
Solution: Calculate in O(N) timeˆaMR

29. WIPL-D / AntNet Integration
29
Array Layout
Input File
WIPL-D AntNet
ˆVSAlg
∝ ˆaAlg ∈ CN×1
Beamformed
Fields
[1] D.S. Weile and D.S. Linden, “AntNet: A fast network analysis add-on for
WIPL-D, 27th International Review of Progress in Applied Computational
Electromagnetics, March 2011
ˆVSnom
= 1 ∈ ℜN×1
O(N3) O(N)
Simulate Once Run Multiple Times
(Saved in file)
Chosen by
Algorithm
Nominal Port
Far Fields &
S/Y/Z matrices

30. HFSS and MOM Models of Array
30
• HFSS = High Frequency Structure Simulator
• HFSS is based on the Finite Element Method (FEM)
• Divides structure into small tetrahedra with boundary conditions
• MOM: divides wires into small segments, planes into small triangles
FEM (HFSS) MOM (WIPL-D)

31. Comparison of HFSS and WIPL-D to in-Situ
Measurements
31
• Good agreement between simulations
and in-situ measurements
• Good agreement in jammer directions
• Extra nulls via nonlinear hardware
effects not captured by HFSS & MOM
HFSS results similar to WIPL results in both cases

32. Diagnosis Model for Hardware Fault Detection
32
Problem: Events overlap making it insufficient to diagnose what caused the
algorithm to fail in anti-jamming by tracking the fitness function alone.
Solution: Add array weight tracking to understand why the algorithm failed.

33. Diagnosis Model for Hardware Fault Detection
• H0: Algorithm converged: No Faults, no TVDOAs.
• H1: Algorithm unconverged: No Faults, no TVDOAs.
• H2: Algorithm converged: Faults and/or TVDOAs present.
• H3: Algorithm unconverged: Faults and/or TVDOAs present.
33
Good states
Not good
states
Diagnosis not
necessary
Diagnosis not possible
∂µa
∂t
n( )> 0and
∂µF
∂t
n( )> 0 →HW Fault
∂µa
∂t
n( )≤ 0and
∂µF
∂t
n( )> 0 →TVDOA
Assumes fading averaged out

34. Antenna Fault Localization: Array Factor Method
AF θ,φ k( )= ˆaiejβaR⋅di
i=1
k−1
∑ + ˆaiejβaR⋅di
i=k+1
N
∑
ˆai = aie− jψi
ai ∈ ℜ and 0 < ai ≤1
ψi ∈ ℜ and 0 ≤ ψ < 2π
Probability that an antenna fault occurred in branch k:
PFault k Failure( )=
1
ξ
max xcorr ARF θ,φ k( ), ARM θ,φ( )!
"
#
$
ARF θ,φ k( )=EF θ,φ( )⋅ AF θ,φ k( )
Note :ξ = normalizing factor s.t. 0 ≤ PFault k Failure( )≤1 34
Complex array
weights
Array Radiated Fields
Element Factor
Array
Factor
K = argmax
k
PFault k Failure( ){ }Assuming 1 fault, most likely fault branch:

35. Antenna Fault Localization with Array Factor
Multiple antenna fault detection possible by counting number of
faults with more calculations due to possible combinations:
PFault k Failure( )=
1
ξ
max xcorr ARF θ,φ k( ), ARM θ,φ( )!
"
#
$
Note :ξ = normalizing factor s.t. 0 ≤ PFault k Failure( )≤1
Single fault:
K faults: PFault k1,,kK[ ] Failure( )=
1
ξ
max xcorr ARF θ,φ k1,,kK[ ]( ), ARM θ,φ( )!
"
#
$
Total AF Correlations =
N
J
!
"
#
$
%
&
J=1
N−1
∑
# Antennas K Total AF Correlations Total AF Corr., 10% Sparsity
4 14 4
8 254 8
16 65534 136
32 > 4 trillion 41448
Total AF Correlations
with Sparsity S s.t. S• N!" #$≥1
=
N
J
&
'
(
)
*
+
J=1
S•N!" #$
∑
35

36. Antenna Fault Localization: Improvements
Pros:
• O(K) for single fault using array factor (AF)
• Useful for small arrays
Cons:
• Correlation fidelity questionable since AF neglects mutual coupling
– Higher fidelity requires MOM or FEM at O(N3) cost
• Less useful for modeling damaged components (i.e, stuck-at faults)
36
Solution: Replace AF calculations with AntNet post-processed
MOM calculations
[1] D.S. Weile and D.S. Linden, “AntNet: A fast network analysis add-on for
WIPL-D,” in the 27th International Review of Progress in Applied
Computational Electromagnetics, March 2011
ˆV = ˆZ ˆZ + ˆZo( )
−1
ˆVS
Eψ θ,φ k( )= Vi
i=1
k−1
∑ Eψ
i
θ,φ( )+ ViEψ
i
θ,φ( )
i=k+1
N
∑ , ψ ∈ θ,φ{ }
PFault k Failure( )=
1
ξ
max xcorr Eψ θ,φ k( ), ARM θ,φ( )!
"
#
$

37. Outline
• Motivation
• Previous Research
• Research Problems and Solutions
• Research Approach
• Experiments and Results
• Conclusion
37

38. Experiment Setup
38
Port 1 Port 2
VNA
VNA = Vector Network Analyzer
SOI
Jammers
SOI = Signal of Interest
SOIJam 1Jam 2Jam 3

39. Adaptive Beamforming Array
39
Goal: Show that stochastic algorithms can perform anti-
jamming beamforming in the presence of static or mobile
signals and hardware faults
Step AttenuatorsPhase Shifters
Antennas
Hardware
Controllers
Power
Combiner

40. Adaptive Beamforming Array in Anechoic Chamber
40
Anechoic chamber approximates free-space conditions
(at far end of
chamber)

41. Array Diagram and Hardware Settings
41
Att5 (8 dB) Att4 (4 dB) Att3 (2 dB) Att2 (1 dB) Att1 (½ dB)
0 1 0 0 1
Ph5 Ph4 Ph3 Ph2 Ph1
1 1 0 0 1
A21 A31 A41 P2 P3 P4
5 BITS5 BITS5 BITS5 BITS5 BITS5 BITS
BIT 30BIT 15BIT 1
Example: 4.5 dB Attenuation
out of 15.5 dB max
Range: 0 to 360°
11.6° / bit Example: 151°
A22 = 0 dB
A32 = 0 dB
A42 = 0 dB
230 Combinations

42. Multimodal SINR Fitness Landscape
42
Collected from 30 independent in-situ SGA runs with two jammers
Multimodal behavior clear with several peaks having SINR ≥ 30 dB

43. Table of Simulations and
Experiments Performed
Anti-Jamming Fault Recovery (Anti-Jamming)
Static Mobile Static Mobile
Algorithm 2 jam 3 jam 2 jam 2 jam 3 jam 2 jam
SGA S, E S, E S S S S
TDGA S, E S S,E S, E S, E S, E
SA S, E S, E S S S S
HC S S S S S S
43
SGA = Simple Genetic Algorithm
TDGA = Triallelic Diploid GA
SA = Simulated Annealing
HC = Hill Climbing
S = Simulation
E = In-situ experiments
First SecondThird

44. Simulated Anti-Jamming with SGA and TDGA: Two
Static Jammers
44
SGA TDGA
PerformanceHamming
• TDGA produced better converged SINR values than SGA
• TDGA mean-Hamming distance decayed slower than SGA
• Mean SINR with 95% confidence interval indicate average convergence by 15
generations for both SGA and TDGA

45. In-Situ Anti-Jamming with SGA & TDGA: Two Static Jammers
45
SGA TDGA
PerformanceHamming
• In-Situ TDGA produced better minimum converged SINR values than SGA
• Difference between min/max SINR smaller for TDGA than SGA
• Simulated SINR values were conservative compared to in-situ results.

46. SGA and TDGA Radiation Patterns, Simulations and In-Situ
Compared: Two Static Jammers
46
SGATDGA
Simulation In-Situ
Simulation and in-situ radiation patterns are similar at convergence

47. Simulated Anti-Jamming with SA & HCA: Two Static Jammers
47
SA HCA
PerformanceRadiation
• SA and HCA obtain similar converged azimuth radiation plots
• Both SA and HCA by chance find ~20 dB SINR solutions early but on
average converge much slower than GAs per 95% confidence intervals

48. In-Situ Anti-Jamming with SA: Two Static Jammers
48
PerformanceRadiation
• Average convergence time agrees with simulations.
• Final in-situ SINR values higher than SINR predicted by simulation

49. Anti-Jamming Two Static Jammers: Algorithm
Comparison
Best Case SINR
(dB)
In-Situ {Sim}
Worst Case
SINR (dB)
In-Situ {Sim}
95% Conf. Interval (dB)
Gauss (Student-t)
In-Situ {Sim}
Average Converge Time
(# Gen / Eval)
In-Situ {Sim}
SGA 67.8
{27.7}
28.3
{20.7}
3.3 (3.4)
{0.65 (0.68)}
15 Gen (3000 Eval)
{15 Gen (3000 Eval)}
TDGA 55.5
{28.0}
31.1
{22.5}
2.4 (2.5)
{0.55 (0.57)}
30 Gen (6200 Eval)
{15 Gen (3000 Eval)}
SA 48.2
{26.1)
13.8
{21.1}
2.52 (2.62)
{0.51 (0.53)}
7800 Eval (39 Gen)
{7140 Eval (~36 Gen)}
HCA {26.2} {18.2} {0.61 (0.63)} {7140 Eval (~36 Gen)}
49
• Simulations were conservative in predicting final SINR values
• SGA and TDGA converged to higher SINR values than SA and HCA
• In-Situ 95% confidence intervals were higher than predicted by simulations
due to hardware tolerances.
• SA and HCA on average converged slower than SGA and TDGA

50. SGA and TDGA Two Jammer Fault Recovery Performance
and Hamming Distance Plots: Simulations
50
SGA TDGA
PerformanceHamming
• SGA and TDGA simulations predict recovery, but simulations are conservative.
• Mean Hamming distance for TDGA decays slower than SGA

51. TDGA In-Situ Fault Recovery Performance and Hamming
Distance Plots
51
PerformanceHamming
• TDGA in-situ experiments recovered with higher final values than simulations.
• 95% confidence intervals indicate TDGA recovered from a fault

52. SGA and TDGA Fault Recovery Azimuth Plots, Simulations
52
SGA
TDGA
Similar final radiation patterns with conservative fault-recovery predicted.

53. TDGA In-Situ Fault Recovery Azimuth Plot
53
[1] J. Becker, J.D. Lohn, and D. Linden, “Towards a self-healing, anti-jamming
adaptive beamforming array,” in 2013 IEEE-APS Topical Conference on
Antennas and Propagation in Wireless Communications (APWC),
September 2013, pp. 1–4.
• TDGA in-situ pattern showed recovery of anti-jamming function.
• Some SOI gain recovered after the fault.
• Null directed at Jammer 2 (J2) deeper than pre-fault null.

54. Why TDGA Self Heals: An Example
54
Went from High to Low
Mean Population Fitness Down
Long term genetic memory and +1’s
to -1’s dominance allows healing

55. SA and HCA Fault Recovery Performance and Azimuth
Radiation Plots: Simulations
55
SA HCA
PerformanceRadiation
• Temperature schedules repeated 5 times to allow for fault recovery
• Both SA and HCA showed fault recovery with maximum ~20 dB SINR post-fault

56. Two Static Jammer Fault Recovery: Algorithm
Comparison
Best Case SINR
(dB) post-Fault
In-Situ {Sim}
Worst Case SINR
(dB) post-Fault
In-Situ {Sim}
95% Conf. Interval
(dB)
Gauss (Student-t)
In-Situ {Sim}
Average Converge
Time (# Gen / Eval)
In-Situ {Sim}
SGA {18.7} {13.4} {0.48 (0.50)} {15 Gen (3000 Eval)}
TDGA 47.0
{18.6}
1.12
{13.6}
4.77 (6.74)
{0.57 (0.59)}
30 Gen (6200 Eval)
{15 Gen (3000 Eval)}
SA {20.5} {13.9} {0.60 (0.63)} Repeated Cooling
Schedules
HCA {20.3} {15.6} {0.43 (0.45) } Repeated Cooling
Schedules
56
Fault Condition: two step attenuators in one path set to full values.
• SGA and TDGA simulations produced similar post-fault SINR values
• SA and HCA simulations produced slightly better SINR results than GA
• Algorithm simulations produced similar 95% confidence intervals but TDGA in-
situ 95% confidence intervals much larger due to hardware tolerances

57. SGA Tracking Two Jammers from {45°, 200°} to {120°, 300°}
57
• SGA moves nulls to track the jammers.
• Previous solution sometimes repeated resulting in lower SINR fitness.

58. TDGA Tracking Two Jammers from {45°, 200°} to {120°, 300°}
58
TDGA behaves in fashion similar to SGA.

59. SGA and TDGA Two Mobile Jammers Constantly Moving:
Simulations
59
SGA TDGA
PerformanceHamming
• SGA and TDGA performance graph follow similar sinusoidal pattern.
• TDGA mean-Hamming distance higher than SGA indicating more diversity in
TDGA populations.

60. Azimuth Radiation Plots for SGA and TDGA Two Mobile
Jammers Constantly Moving: Simulations
60
SGA
TDGA
Both SGA and TDGA track both jammers with second
jammer having deeper null.

61. Stochastic Algorithms Investigated
Name Advantages Disadvantages
Simple Genetic
Algorithm (SGA)
Able to search parameter
space in parallel
Complexity problem
dependent, short-term
genetic memory
Triallelic Diploid
Genetic Algorithm
(TDGA)
Able to search parameter
space in parallel, long-term
genetic memory
Complexity problem
dependent, added step to
convert TD strings into
binary haploid strings
Simulated Annealing
(SA)
Temperature dependent
mutation allows initial
exploration of search
space with eventual
exploitation of solutions
Convergence time
temperature schedule
dependent, 2X slower
than GAs
Hill Climbing
Algorithm (HCA)
Simple to implement, finds
solutions comparable to
GAs and SA
Tends to get stuck at local
optima, 2X slower than
GAs
61

62. Results Discussion
• Incorporating physical objects into MOM model of array increased model
reliability and fidelity compared to in-situ measurements.
• Need to track both fitness function values and complex weights for a
useful diagnostic model to detect faults in non-ideal environments.
• Hardware faults can be localized by correlating in-situ measurements with
MOM calculations to provide most-likely faulty antenna branch.
• Showed that stochastic algorithms can perform anti-jamming
beamforming with hardware fault recovery
– Simulations gave conservative results in SINR values compared to in-situ
measurements
– Simulated Annealing and Hill Climbing Algorithms slower than GAs at anti-jamming
static signals.
• GAs able to thwart continuously moving jammers
62

63. Conclusions and Contributions
• New analytical models with experimental results showing that small
antenna arrays can thwart interference sources with unknown positions.
• First time demonstration of in-situ optimization with an algorithm
dynamically optimizing a beamformer to thwart interference sources with
unknown positions, inside of an anechoic chamber.
• First time demonstration of stochastic algorithms that provided recovery
from hardware failures and localized faults in the array with reconfiguration
of array weights to provide anti-jamming of interference sources having
unknown positions.
• Comparison of multiple stochastic algorithms in performing both anti-
jamming and hardware fault recovery.
• Showed that stochastic algorithms can be used to continuously track and
mitigate interfering signals that continuously move in an additive white
Gaussian noise wireless channel.
63

64. Future Work
• Real-time fault recovery and anti-jamming in wireless link
• Wideband 8-antenna array with individual antenna modules
64
PN = Pseudo-random Noise
USRP = Universal Software
Radio Protocol

65. Selected Publications
1. J. Lohn, J. M. Becker, and D. Linden, “An evolved anti-jamming adaptive beam-forming
network,” Genetic Programming and Evolvable Machines, vol. 12, no. 3, pp. 217–234, 2011.
2. J. Becker, J. Lohn, and D. Linden, “An anti-jamming beamformer optimized using evolvable
hardware,” in Proc. 2011 IEEE Intl. Conf. on Microwaves, Communications, Antennas, and
Electronic Systems, IEEE COMCAS 2011, November 2011, pp. 1–5.
3. J. Becker, J. D. Lohn, and D. Linden, “An in-situ optimized anti-jamming beamformer for
mobile signals,” in 2012 IEEE International Symposium on Antennas and Propagation, IEEE
APS 2012, July 2012, pp. 1–2.
4. J. Becker, J. Lohn, and D. Linden, “Evaluation of genetic algorithms in mitigating wireless
interference in situ at 2.4 GHz,” in WiOpt 2013 Indoor and Outdoor Small Cells Workshop,
May 2013, pp. 1–8.
5. J. Becker, J. D. Lohn, and D. Linden, “Algorithm comparison for in-situ beamforming,” in
2013 IEEE Intl. Symp. on Antennas and Propagation, IEEE APS 2013, July 2013, pp. 1–2.
6. J. Becker, J. D. Lohn, and D. Linden, “Towards a self-healing, anti-jamming adaptive
beamforming array,” in 2013 IEEE-APS Topical Conference on Antennas and Propagation
in Wireless Communications (APWC), September 2013, pp. 1–4.
65

66. Acknowledgements
• I thank my committee members for their support:
– Professor Jason Lohn
– Professor Ole Mengshoel
– Professor Patrick Tague
– Dr. Derek Linden, CTO X5 Systems Inc.
• I would also like to thank these individuals who assisted me over the years:
Prof. Martin Griss, Prof. Bob Iannuchi, Prof. Ted Selker, Dr. James Downey,
Dr. Reggie Cooper, Prof. Joshua Griffin, Dr. Matthew Trotter, Prof. Joy Zhang,
Prof. Pei Zhang, Prof. Emeritus James Hoburg, Prof. James Bain, Dr. Joey
Fernandez, Dr. Faisal Luqman, Dr. Heng-Tze Cheng, Dr. Joel Harley, Jon
Smereka
• This research was funded in part by:
– Cylab at Carnegie Mellon University under grant DAAD19-02-1-0389 from the Army Research
Office
– The Electrical and Computer Engineering Department at Carnegie Mellon University
66

67. Thank you
67