Optimizing Systems with Conﬂicting Objectives Competing

Optimizing Systems with Conflicting ObjectivesCompeting for a Limited Resource

Radar waveform design, Unimodular QPUAV tracking optimization

Hans D. Mittelmann

School of Mathematical and Statistical SciencesArizona State University

AFOSR Optimization and Discrete Math Review

23 August 2019

COLRO problems Hans D. Mittelmann MATHEMATICS AND STATISTICS 1 / 36

collaborators

Shankarachary Ragi, South Dakota School of Mines & Technology

Daniel Bliss, Alex Chiriyath, Arizona State University

Edwin Chong, Colorado State University

Shawon Dey, Azam Md Ali, South Dakota School of Mines & Technology


Outline

Waveform Design for Joint Radar-CommunicationsBackgroundWaveform Optimization MethodsNumerical Results

Unimodular Quadratic ProgramsBackgroundProblem DescriptionExisting MethodsOur MethodsNumerical Results

Ongoing ResearchCompeting Objective Optimization in Networked Swarm Systems


Introduction

• Traditionally, wireless communications (0.3 - 3 GHz) and radar (3 - 30GHz) are spectrally separated

• Spectral congestion forcing co-existence

• Key performance factors: spectral shape of waveform, receiver design,signal decoupling strategies


Radar: Preliminaries

Radar Transmitter

Radar Target

Transmitted signal

Waveform transmission → Scattered signal recovery → Matched filter response

• Signal delay→ Range detection• Doppler shift→ Speed detection


Joint Radar-Comms System

Radar Target

Joint

Radar-Communications

Node

Communications

Transmitter

Radar Spectrum

Communications Spectrum

Key functions:• Sends radar pulses for target detection

• Receives a mixed signal - radar return and communications signal

• Decouples the signals

• Processes radar returns for ranging and speed


Signal Decoupling

Successive-Interference Cancellation

TransmitRadar

Waveform

RadarChannel

CommsChannel

RemovePredicted

Return

DecodeComms

& Remove

ProcessRadarReturn

Comms InfoCommsSignal

Σ

Key step: Remove predicted radar return from the mixture


Performance Indicators

• Communications performance: Shannon’s information rate bound

Rcom ≤ B log2

(1 +‖b‖2 Pcom

σ2int+n

)

• Radar performance: estimation rate bound [D. W. Bliss, 2014 IEEE RadarConference]

Rest ≤δ

2Tlog2

[1 +

σ2proc

σ2est

]

Spectral shape of the waveform directly influences the above per-formance indicators!

σ2int+n ∝ Brms σ2

est ∝1

Brms


Spectral-Shape Affects Performance

−B/2 0 B/2

Sp

ectr

al

Mask

Weig

hti

ng

FrequencySNR (dB)

Radar optimal

Comms optimal


Maximizing joint radar-communications performanceu controls the spectral-shape of the waveform!

maximizeu∈[0,1]N

[Rcom(u)]α [Rest(u)]1−α

subject to system constraints

Result: If u∗ is the optimal solution, then u∗ is pareto efficient[S. Ragi, A. Chiriyath, D. Bliss, H. Mittelmann, Optimization-Online pre-print]

Estimationrate

Commsrate


Spectrum: α = 0 and α = 1

-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5Frequency (Hz) 10 6

0

2

4

6

8

10

12

14

16

18

20A

mpl

itude

Frequency spectrum vs. Emphasis on radar( = 0)Emphasis on comms( = 1)


Solver Performance

0 500 1000 1500Rest

4.5

5

5.5

6

6.5

7

7.5

8

8.5

9

9.5

10R

com

10 6 Rest vs. R comm

KnitroFminconDE with 500 Iter

Increasing

= 1

= 0

[S. Ragi, A. Chiriyath, D. Bliss, H. Mittelmann, Optimization-Online pre-print]


Solver Performance

10 -1 10 0 10 1 10 2 10 3

Execution time (sec, logscale)

0

0.2

0.4

0.6

0.8

1C

umul

ativ

e fr

eque

ncy

Empirical CDF

EvolFminconKnitro


Outline





Monostatic Radar

• Transmits encoded pulse sequence

a(0)

a(1) a(2) a(3)

time

u(t)

• Objective: optimize c = (a(0), . . . ,a(N))T, where |a(i)| = 1 ∀i , tomaximize signal-to-noise ratio (SNR)

SNR = cHRc

where R = M−1 (ppH)∗, M = E[wwH]

• Code optimization leads to unimodular quadratic program (UQP)


Unimodular Quadratic Program• Ω = x ∈ C, |x | = 1

1-1

-i

iΩ

• Unimodular quadratic program (UQP)

maximizes∈ΩN

sHRs

where R ∈ CN×N is a Hermitian matrix.

• Many problems in radar and wireless communications lead to UQP• UQP is an NP-hard problem

[S. Zhang, et. al., “Complex quadratic optimization and semidefinite programming," SIAM J. Optimization,2006]


Semi-Definite Relaxation (SDR)

• UQP can also be stated as (since sHRs = tr(sHRs) = tr(RssH))

maximizeS

tr(RS)

subject to S = ssH , s ∈ ΩN .

• If rank constraint is relaxed⇒ semidefinite program (SDP)

maximizeS

tr(RS)

subject to [S]k,k = 1, k = 1, . . . ,NS is positive semidefinite.

• SDP can be solved in polynomial time


Phase-Matching with Dominant Eigenvector

• Pick d ∈ ΩN that “phase-matches” the dominant eigenvector of R

• Time complexity: O(N3)

• Example:

If (0.2eiπ/3, 0.6eiπ/4, 0.77eiπ/5)T is the dominant eigenvector, thend = (eiπ/3, eiπ/4, eiπ/5)T

[S. Ragi, E. K. P. Chong, H. D. Mittelmann, “Heuristic methods for designing unimodular code sequences withperformance guarantees,” ICASSP 2017.]


Performance Bound

Result (S. Ragi, E. K. P. Chong, H. D. Mittelmann, ICASSP 2017)If VD = dHRd and Vopt is the optimal objective value, then

VDVopt

≥ λN + (N − 1)λ1

λNN

N λ1 λN Bound34 6.7 81.4 0.1196 8.7 64.6 0.1493 19.5 71.6 0.286 41.6 50.8 0.8574 40.2 99.2 0.4127 3 58.4 0.09


Greedy Strategy

• Greedy solution g = (g(1), . . . ,g(N))T

g(k) = arg maxx∈Ω

[gk−1; x ]HRk [gk−1; x ],

k = 2, . . . ,N, g(1) = 1

gk = (g(1), . . . ,g(k))T

where Rk is the k × k principle sub-matrix of R.

Example: If R =

1 2 32 4 53 5 6

, then R2 =

[1 22 4

], and R3 = R


Performance Bound for Greedy Strategy

If the objective function is string submodular, then

gHRg ≥ (1− 1/e) maxs∈ΩN

sHRs

where (1− 1/e) ≈ 0.63


String Submodular FunctionsMonotonic functions with diminishing returns!

Input

Output


Is UQP objective function string submodular?

f (Ak ) = AHk Rk Ak

• f is not string-submodular⇒ UQP for any R is not string-submodular

• But f (Ak ) = AHk Rk Ak is string submodular, where R is obtained from R

via manipulating diagonal entries of R

[S. Ragi, E. K. P. Chong, H. D. Mittelmann, “Heuristic methods for designing unimodular code sequences withperformance guarantees,” ICASSP 2017.]


Bound for greedy method

Result (S. Ragi, E. K. P. Chong, H. D. Mittelmann, ICASSP 2017)If Tr(R) ≤ Tr(R), then

gHRg ≥(

1− 1e

)(maxs∈ΩN

sHRs),

where g is the solution from the greedy method.

• Time complexity of greedy method: O(N)


Performance Comparison

0 20 40 600

0.2

0.4

0.6

0.8

1

F(x

)

Obj Val N=10GreedyDom. Eig.Row-SwapGreedySDRPowerMethod

0 50 100 150

Objective value

0

0.2

0.4

0.6

0.8

1Obj Val N=20

0 100 200 3000

0.2

0.4

0.6

0.8

1Obj Val N=30

10 -5 10 00

0.2

0.4

0.6

0.8

1

F(x

)

Exec Time N=10

10 -4 10 -2 10 0 10 2

Execution time (sec, logscale)

0

0.2

0.4

0.6

0.8

1Exec Time N=20

10 -5 10 00

0.2

0.4

0.6

0.8

1Exec Time N=30


Outline





COLRO Problems

• COLRO: competing objective limited resource optimization

• COLRO problems appear naturally in many applications includingdecision making in autonomous systems

• We explore novel methods to solve COLRO problems in real-time


UAV swarm control• Goal: control the motion of a networked swarm of UAVs while tracking a

target• Minimizing the energy costs and maximizing the tracking performance

are conflicting objectives

Target Target’s trajectory

Swarm centroid


COLRO formulation

• Goal: optimize the motion of UAVs to maximize target trackingperformance while minimizing the network energy costs

• Decision variables: swarm centroid Ck and Gk

minGk ,Ck ,k=0,..,H−1

H−1∑k=0

E[wftrack (Gk ,Ck , χk )+

(1− w)fenergy (Gk ,Ck , χk )]

(1)

• Objective function is hard to evaluate exactly!


COLRO cost functions


H−1∑k=0

E[wftrack (Gk ,Ck , χk )+

(1− w)fenergy (Gk ,Ck , χk )]

(2)

• ftrack measuresI benefits of data fusion between a pair of UAVs - GkI benefits of having the swarm staying close to the target - Ck

• fenergy measuresI benefits of using the communications network sparingly - Gk


Solution Approach

• Nominal belief-state optimization - an approximate dynamic programmingapproachI Replace future noise variables with “nominal” valuesI Replace the expectation with “nominal” trajectory of the posterior distribution

into the future

• Apply receding horizon control approach


H−1∑k=0

[wftrack (Gk ,Ck , ψk )+

(1− w)fenergy (Gk ,Ck , ψk )]

(3)

where ftrack and fenergy are deterministic approximations.

• Mixed integer nonlinear program - solution is obtained via a commercialsolver Knitro


Converting G∗k and C∗k to UAV kinematic controls

Desired centroid

Repulsive force to avoid collisions

Attractive force to form desired network

Attractive force toward centroid


3 UAVs and 1 target (H = 6)

-1000 -500 0 500 1000 1500 2000 2500 3000

0

500

1000

1500

2000

Target trajectory

UAV trajectory


Performance against centralized approach (3 UAVs)

0 50 100 150

Time (sec)

0

5

10

15

20

25

30

35

40

45

Tar

get l

ocat

ion

erro

r (m

)

Target tracking performance vs. centralized approach

centralized fusionapproachsensor 1sensor 2sensor 3


Future Work

• Incorporate “belief consensus” into the COLRO frameworkI Running consensus algorithms leads to increased network energy costs, but

improves cooperativeness of the agentsI Belief consensus can be time consuming - we will develop fast heuristic

approaches

• Dealing with heterogeneous data from sensors on-board the agents, e.g.,imagery, video, and audio. We need new data fusion techniques, e.g.,fusion in feature space


Thank you for your attention!

For papers see http://plato.asu.edu/papers.html no.s 142-152


Documents

Optimizing Systems with Conﬂicting Objectives Competing