UNDERSTANDING CONFLICT-RESOLUTION TASKLOAD:

IMPLEMENTING ADVISORY CONFLICT-DETECTION AND

RESOLUTION ALGORITHMS IN AN AIRSPACE

A Thesis
Presented to
The Academic Faculty

by

Adan Ernesto Vela

In Partial Fulfillment
of the Requirements for the Degree
Doctor of Philosophy in the
School of Mechanical Engineering

Georgia Institute of Technology

December 2011
 UNDERSTANDING CONFLICT-RESOLUTION TASKLOAD:
IMPLEMENTING ADVISORY CONFLICT-DETECTION AND
RESOLUTION ALGORITHMS IN AN AIRSPACE

Approved by:

Prof. William Singhose, Prof. Karen Feigh

Committee Chair School of Aerospace Engineering
School of Mechanical Engineering Georgia Institute of Technology
Georgia Institute of Technology

Prof. John-Paul Clarke Prof. Wayne Book

School of Aerospace Engineering School of Mechanical Engineering
Georgia Institute of Technology Georgia Institute of Technology

Prof. Eric Feron Prof. Kok-Meng Lee

School of Aerospace Engineering School of Mechanical Engineering
Georgia Institute of Technology Georgia Institute of Technology

Date Approved: December 2011

To my loving and supporting family.

iii
 ACKNOWLEDGEMENTS

It is impossible to find the perfect words to thank my advisors, family, colleagues and friends

for supporting my dream of obtaining a Ph.D. in engineering. They have influenced me with

their academic and life wisdom to make me a better engineer and person. In particular,

I want to thank my advisors Dr. William Singhose, Dr. John-Paul Clarke and Dr. Eric

Feron. In Dr. William Singhose, I found more than a mechanical engineering advisor. He

introduced me to Georgia Tech as an undergraduate, and later welcomed me into his lab.

He has always guided my Ph.D. studies with challenging and intriguing questions. Through

Dr. John-Paul Clark and Dr. Eric Feron, I learned to appreciate the practical and the

theoretical sides of air transportation research. Each has taught me to push the boundaries

of research and to never be apprehensive about exploration. I would not have acquired the

knowledge, nor the skills, to complete my Ph.D. studies without the wisdom and guidance

of my advisors.

I also want to thank Dr. Karen Feigh. She taught me the importance of human-

centric air traffic control. I must also thank Dr. Daniel Delahaye for inviting me to study

air traffic control at ENAC. He always encouraged a fun and challenging discourse on air

traffic control. For supporting me from my first days as a Ph.D. student until now, I would

like to thank Dr. Senay Solak. He continues to inspire me with his work in Operations

Research. He will remain a role model and a source of inspiration. I also thank Dr. Senay

Solak for his unwavering encouragement.

In my mechanical engineering lab, I want to give a special thanks and acknowledgement

to Dr. Khalid Sorensen and Dr. Joshua Vaughan, who gave me advice at every turn of my

Mechanical Engineering Ph.D. process and beyond. For teaching me about the realities of

aircraft and air traffic control, I want to acknowledge Jim Brooks and Gaurav Nagel.

Next, I want to acknowledge all my friends and co-workers who have shared and sup-

ported my struggles. A special mention goes to Dr. Kuemjin Lee, Dr. Maxime Gariel, and

iv
Dr. Erwan Saluan, who have been comrades in research.

To conclude, I want to thank my family. They are the greatest support system in my

life, starting with my parents who devoted their lives to inspire me and many young people

to study math, science and engineering. I also thanks my brother and sister who have given

me their love and support throughout my whole life.

v
 SUMMARY

From 2010 to 2030, the number of instrument flight rules aircraft operations handled

by Federal Aviation Administration en route traffic centers is predicted to increase from

approximately 39 million flights to 64 million flights. The projected growth in air trans-

portation demand is likely to result in traffic levels that exceed the abilities of the unaided

air traffic controller in managing, separating, and providing services to aircraft. Conse-

quently, the Federal Aviation Administration, and other air navigation service providers

around the world, are making several efforts to improve the capacity and throughput of

existing airspaces. Ultimately, the stated goal of the Federal Aviation Administration is to

triple the available capacity of the National Airspace System by 2025.

In an effort to satisfy air traffic demand through the increase of airspace capacity, air

navigation service providers are considering the inclusion of advisory conflict-detection and

resolution systems. In a human-in-the-loop framework, advisory conflict-detection and res-

olution decision-support tools identify potential conflicts and propose resolution commands

for the air traffic controller to verify and issue to aircraft. A number of researchers and air

navigation service providers hypothesize that the inclusion of combined conflict-detection

and resolution tools into air traffic control systems will reduce or transform controller work-

load and enable the required increases in airspace capacity.

In an effort to understand the potential workload implications of introducing advisory

conflict-detection and resolution tools, this thesis provides a detailed study of the con-

flict event process and the implementation of conflict-detection and resolution algorithms.

Specifically, the research presented here examines a metric of controller taskload: how many

resolution commands an air traffic controller issues under the guidance of a conflict-detection

and resolution decision-support tool. The goal of the research is to understand how the for-

mulation, capabilities, and implementation of conflict-detection and resolution tools affect

the controller taskload (system demands) associated with the conflict-resolution process,

vi
and implicitly the controller workload (physical and psychological demands). Furthermore

this thesis seeks to establish best practices for the design of future conflict-detection and

resolution systems.

To generalize conclusions on the conflict-resolution taskload and best design practices of

conflict-detection and resolution systems, this thesis focuses on abstracting and parameteriz-

ing the behaviors and capabilities of the advisory tools. Ideally, this abstraction of advisory

decision-support tools serves as an alternative to exhaustively designing tools, implementing

them in high-fidelity simulations, and analyzing their conflict-resolution taskload. Such an

approach of simulating specific conflict-detection and resolution systems limits the type of

conclusions that can be drawn concerning the design of more generic algorithms.

In the process of understanding conflict-detection and resolution systems, evidence in

the thesis reveals that the most effective approach to reducing conflict-resolution taskload

is to improve conflict-detection systems. Furthermore, studies in the this thesis indicate

that there is significant flexibility in the design of conflict-resolution algorithms.

vii
 TABLE OF CONTENTS

DEDICATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii

ACKNOWLEDGEMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv

SUMMARY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi

LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi

LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xii

GLOSSARY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvi

I INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Thesis Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.2 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

II AIRSPACE MODEL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.1 Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2 Data Set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3 Sampling Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.4 Separation Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.5 Application and Verification of the Sampling Procedure . . . . . . . . . . . 27
2.5.1 Spatial Distribution of Aircraft . . . . . . . . . . . . . . . . . . . . 27
2.5.2 Maximum Altitude Distribution of Aircraft . . . . . . . . . . . . . 29
2.5.3 Distribution of Aircraft Models . . . . . . . . . . . . . . . . . . . . 30
2.6 Scaling for Equitable Comparison . . . . . . . . . . . . . . . . . . . . . . . 31
2.7 Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

III SECTOR CONFLICT ANALYSIS . . . . . . . . . . . . . . . . . . . . . . . 38

3.1 Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.2 Uncontrolled Conflicts within a Sector . . . . . . . . . . . . . . . . . . . . 43
3.3 Controller Communication and Control Time . . . . . . . . . . . . . . . . 59
3.4 Review and Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

IV GRAPH REPRESENTATIONS OF POTENTIAL CONFLICTS . . . 68

4.1 Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

viii
 4.2 Representing aircraft and potential conflicts through graphs . . . . . . . . 72
4.3 Review of Conflict Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

V ABSTRACTION OF CONFLICT-DETECTION AND RESOLUTION

TOOLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.1 Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.2 Receding-Horizon Control . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.3 Parameterization of Conflict-Detection Systems . . . . . . . . . . . . . . . 85
5.4 Parameterization of Conflict-Resolution Systems . . . . . . . . . . . . . . . 88
5.4.1 Implementations and Capabilities: HD , HR , and δt . . . . . . . . . 89
5.4.2 Conflict-Resolution Policies . . . . . . . . . . . . . . . . . . . . . . 94
5.5 Solution Times . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.6 Special Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.7 Clearance-Based Control versus Trajectory-Based Operations . . . . . . . 105
5.8 Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

VI SIMULATION MODELING FOR ABSTRACTED TOOLS . . . . . . 110

6.1 Creation of Complete Conflict-Graphs . . . . . . . . . . . . . . . . . . . . 111
6.2 Simulation Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
6.3 Parameter Settings and Traffic Scenarios . . . . . . . . . . . . . . . . . . . 114

VII SIMULATION ANALYSIS . . . . . . . . . . . . . . . . . . . . . . . . . . . 116

7.1 Aggregate Behaviors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
7.2 Initial Parameter Sensitivity Analysis . . . . . . . . . . . . . . . . . . . . . 122
7.3 Policy Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
7.4 min and H
Relative value between Dsep R . . . . . . . . . . . . . . . . . . . . . 132
7.5 Benefit of Improved Conflict-Resolution . . . . . . . . . . . . . . . . . . . . 139
7.6 Managing Uncertainty and Solve-Times . . . . . . . . . . . . . . . . . . . . 142
7.7 Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147

VIII CONCLUSION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150

8.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
8.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
8.3 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153

ix
APPENDIX A — DECISION POLICIES . . . . . . . . . . . . . . . . . . . 157

REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163

x
 LIST OF TABLES

1 Best-fit coefficients for the uncontrolled conflict models. . . . . . . . . . . . 50

2 Coefficients and R2 values for least-squares lines of best-fit considering the
ratio of conflict complexity. . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3 Average communication time for each event. . . . . . . . . . . . . . . . . . . 62
4 Worst-case model errors between simulation averages and predicted averages
over all configurations, traffic intensities, and sectors. . . . . . . . . . . . . . 118
5 Policy Comparisons: ZMP12 . . . . . . . . . . . . . . . . . . . . . . . . . . 127
6 Policy Comparisons: ZMP16 . . . . . . . . . . . . . . . . . . . . . . . . . . 127
7 Policy Comparisons: ZMP42 . . . . . . . . . . . . . . . . . . . . . . . . . . 127
8 Configurations with Dsep min ≤ 5N M for which there is a statistical difference

in the conflict-resolution taskload distribution over 2.5 minutes for ZMP42 . 131
9 Knowledge-Based Policy Rules . . . . . . . . . . . . . . . . . . . . . . . . . 160

xi
 LIST OF FIGURES

1 Map of the 20 Air Route Traffic Control Centers covering the continental
United States. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2 Center map for Minneapolis Center (ZMP) with high-alitude sector boundaries. 3
3 Under current operations controllers use radar and vocal communication sys-
tems to track and issue commands to aircraft. . . . . . . . . . . . . . . . . . 5
4 Conflict-detection and resolution decision-support tools provide possible res-
olution solutions to air traffic controllers. . . . . . . . . . . . . . . . . . . . 7
5 Representation of factors affecting controller workload.[66] . . . . . . . . . . 8
6 Chapter dependency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
7 High-altitude traffic pattern for June 14th , 2007 in Minneapolis Center. . . 18
8 High-altitude boundaries for Areas in ZMP. . . . . . . . . . . . . . . . . . . 18
9 High-altitude sectors boundaries in Area 6 of ZMP. . . . . . . . . . . . . . . 19
10 Diurnal aircraft counts within Minneapolis Center for the dates May, 21,
2007 through June 17th, 2007. . . . . . . . . . . . . . . . . . . . . . . . . . 20
11 Spatial clustering of origins described by latitude and longitude. . . . . . . 21
12 Distribution for origin-destination pairs for flights within Minneapolis Center. 22
13 Clustering relationships between repeated origin-destination pairs and air-
craft fights. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
14 Example conflict between two aircraft. . . . . . . . . . . . . . . . . . . . . . 26
15 Spatial distribution of aircraft for June 14th , 2007 and a generated traffic
scenario. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
16 Distribution of maximum aircraft altitudes for June 14th , 2007 and a gener-
ated traffic scenario. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
17 Distribution of aircraft models for June 14th , 2007 and a generated traffic
scenario. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
18 Illustrative example of the subsampling procedure. . . . . . . . . . . . . . . 33
19 Extrapolated conflict totals based on sub-sampled traffic scenarios. . . . . . 34
20 Coefficient of determination, R2 for models based on sub-sampled traffic
scenarios used to predict conflict totals as a function of traffic intensity in
each sector. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
21 Example of conflict graph G with three conflict-clusters. . . . . . . . . . . . 40
22 Potential conflicts can be characterized by the number of aircraft involved
and overall complexity of generating a resolution. . . . . . . . . . . . . . . . 41

xii
23 Conflict totals for sectors in Minneapolis Center over a 24hr time period when
sep
Dmin = 9 NM. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
24 Traffic Density Maps. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
25 ZMP12 : Probability distribution of flight phases. . . . . . . . . . . . . . . . 46
26 ZMP42 : Probability distribution of flight phases. . . . . . . . . . . . . . . . 46
27 ZMP16 : Probability distribution of flight phases. . . . . . . . . . . . . . . . 46
28 Location of uncontrolled conflicts according to conflict configuration. . . . . 47
29 Discrete empirical probability distribution of uncontrolled conflict crossing
angles (5o increments). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
30 Cumulative number of uncontrolled conflicts for sectors ZMP12, ZMP17, and
sep
ZMP42, when Dmin = 9 NM. . . . . . . . . . . . . . . . . . . . . . . . . . . 49
31 Cumulative number of aircraft arrivals for sectors ZMP12, ZMP17, and ZMP42. 49
32 Simulated number of uncontrolled conflicts and best-fit lines according to
traffic intensity and separation distance for ZMP42. . . . . . . . . . . . . . 51
33 min = 9 NM. . . . . . . . . . .
Average rate of uncontrolled conflicts when Dsep 52
34 Cumulative distribution of the number of uncontrolled conflicts for the bus-
min = 9 NM for ZMP12. . . . . . . .
iest 5 minute period of the day when Dsep 53
35 Cumulative distribution of the number of uncontrolled conflicts for the bus-
min = 9 NM for ZMP16. . . . . . . .
iest 5 minute period of the day when Dsep 53
36 Distribution of the number of uncontrolled conflicts for the busiest 5 minute
min = 9 NM for ZMP42. . . . . . . . . . . . . . .
period of the day when Dsep 54
37 Distribution of the number of aircraft invovled in each conflict cluster as
min = 9 NM for ZMP42. . . . . . . . . . . . .
traffic intensity increases for Dsep 55
38 Distribution of the number of edges in each conflict cluster as traffic intensity
min = 9 NM for ZMP42. . . . . . . . . . . . . . . . . . . . .
increases for Dsep 55
39 Fraction of pairwise uncontrolled conflicts. . . . . . . . . . . . . . . . . . . . 56
40 Average growth of conflict clusters as intensity increases. . . . . . . . . . . . 57
41 Number of expected conflicts versus number of expected aircraft in each cluster. 58
42 Two conflict clusters of differing conflict complexity. . . . . . . . . . . . . . 58
43 Representation of the event process within an airspace, and method for cal-
culating the minimum estimated controller communication time. . . . . . . 61
44 Average utilization time with upper and lower quartiles for ZMP42. . . . . 63
45 Average utilization time with upper and lower quartiles for Tw = 5 minutes. 63
46 Probability of over-utilization for different time-windows, Tw . . . . . . . . . 65

xiii
47 Air traffic controller deferring to an advisory conflict-detection and resolution
system. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
48 Example graph representation of aircraft and potential conflict relationships. 73
49 Conflict-resolution problem solved in a receding-horizon control framework. 84
50 Safety region around an aircraft (not to scale) . . . . . . . . . . . . . . . . . 85
51 Parameters determining labeling of potential conflicts. . . . . . . . . . . . . 86
52 Representation of potential conflicts that are considered in the decision-
making process. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
53 Larger values of the guaranteed conflict-free time, HR , help to prevent sec-
ondary conflicts. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
54 With solve-times of δt, trajectory predictions are updated for each imple-
mentation of the conflict-resolution tool. . . . . . . . . . . . . . . . . . . . . 94
55 Sector map of ZMP42 with jetroutes . . . . . . . . . . . . . . . . . . . . . . 98
56 Example application of the minimum vertex cover problem. . . . . . . . . . 100
57 In the planar case, there exist two topologically distinct options for each
resolution command. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
58 Finding solution to the motion planning problem becomes increasingly diffi-
cult as time horizons increase. . . . . . . . . . . . . . . . . . . . . . . . . . . 104
59 Two example resolution solutions for a potential conflict . . . . . . . . . . . 106
60 Difference between avoidance maneuvers and closed-loop resolution commands.107
61 Traffic density map of ZMP center, highlighting sectors of interest. . . . . . 115
62 Qualitative behavior of conflict-resolution taskload with regards to the traffic
intensity and aircraft spacing. . . . . . . . . . . . . . . . . . . . . . . . . . . 118
63 Qualitative behavior of conflict-resolution taskload with regards to the un-
certainty and solve-time. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
64 Qualitative behavior of conflict-resolution taskload with regards to the decision-
horizon time. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
65 Qualitative behavior of conflict-resolution taskload with regards to conflict-
free resolution time. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
66 FCFS parameter sensitivity analysis: Range of percent reduction in conflict-
resolution taskload. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
67 MCRT parameter sensitivity analysis: Range of percent reduction in conflict-
resolution taskload. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
68 Representative timeline of conflict-resolution advisories. . . . . . . . . . . . 126

xiv
69 Percent of configurations for which the MCRT policy is statistical different
from a rule-based heuristic policy (Reject Ho ). . . . . . . . . . . . . . . . . 128
70 Maximum difference in the expected conflict-resolution taskload between the
MCRT policy and the random policy over a range of configurations. . . . . 129
71 Probability distributions of the conflict-resolution taskload for the busiest 2.5
minute time period in each traffic scenario. . . . . . . . . . . . . . . . . . . 130
72 Expected number of advisory resolution commands for the FCFS policy at
1X, 2X, and 3X traffic intensities. . . . . . . . . . . . . . . . . . . . . . . . . 134
73 Required value of HR to match a reduction in aircraft spacing (ZMP12). . . 136
74 Required value of HR to match a reduction in aircraft spacing (ZMP42). . . 136
75 Required value of HR to be equivalent to a reduction in the separation re-
min = 0.5 NM (ZMP42). . . . . . . . . . . . . . . . . . . .
quirements of ∆Dsep 137
76 Percent decrease in the conflict-resolution taskload by reducing aircraft spac-
ing when compared to Dsep initial = 9 NM. (At the 1X, 2X, and 3X traffic

intensities) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
77 r = 0). . . . . . .
Conflict-resolution taskload for ZMP42 (FCFS, δt = 0, Dsep 139
78 Percent reduction in conflict-resolution taskload compared to HR = 6 min-
utes (ZMP42). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
79 Relative reductions in conflict-resolution taskload for ZMP12 and ZMP16. . 141
80 Distribution of aircraft transit times [minutes] through sector (1 minute in-
crements). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
81 Distribution of inter-arrival times [minutes] between the first and last poten-
tial conflict of an aircraft (1 minute increments). . . . . . . . . . . . . . . . 142
82 Conflict-resolution taskload as a function of the solve-time and uncertainty
parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
83 Required value of HR to match faster solve-times. . . . . . . . . . . . . . . . 146
84 Representative summary of analysis. . . . . . . . . . . . . . . . . . . . . . . 147
85 Taking full advantage of conflict-resolution systems may require a system-
wide airspace redesign. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
86 Different phases of flight considered in the knowledge-based policy. . . . . . 160

xv
 GLOSSARY

Acknowledgement When an aircraft enters an airspace, the pilot must announce his pres-
ence to the air traffic controller managing it. The air traffic controller responds
by acknowledging the pilot.

Air Navigation Service Provider (ANSP) Organization responsible for managing aircraft
in an airspace on behalf of a nation or a number of nations.

Air Route Traffic Control Center (ARTCC) A large volume of airspace, defined by ver-
tical and lateral boundaries. Commonly referred to as centers. Centers are
divided into a number of sectors. In the United States, a center covers several
states.

Air Traffic Controller The acting agent of an air navigation service provider. The air
traffic controller is responsible for managing the traffic and ensuring required
spatial separation between all aircraft in an airspace.

Airspace A generic volume of space defined by vertical and lateral boundaries.

Airspace Capacity The nominal number of aircraft an air traffic controller is able to man-
age under reasonable workload.

Automatic Dependent Surveillance-Broadcast A dedicated data channel and protocol

for communicating between aircraft and ground stations. A major function of
ADS-B is to act as a primary surveillance source for reporting GPS position
information.

Clearance-Based Control The current operational protocol for managing aircraft. Air-
craft trajectories are managed in piecewise increments through a sequence of
request and clearances.

Complexity A generic term relating the difficulty of a particular traffic scenario.

Conflict When any two aircraft break minimum separation requirements. For the en
route environment aircraft must maintain 5NM lateral, and 1,000ft vertical
separation from each other at all times.

Conflict-Resolution Taskload The number of resolution commands used in ensuring sep-

aration between aircraft in a given time period.

Controller Workload The amount of effort, both physical and psychological, expended
in response to system demands (task load) and also in accordance with the
operator’s internal standard of performance.

Direct Routing When aircraft fly between two points in space along the shortest path.

EUROCONTROL The air navigation service provider for much of Europe.

Federal Aviation Adminstration (FAA) The air navigation service provider for the United
States of America.

xvi
Hand-off When an aircraft transitions from one airspace to another, the air traffic con-
troller in the originating airspace provides the aircraft with the communication
frequency of the destination airspace. The pilot of the aircraft is then expected
to communicate with the controller in the destination airspace in order to be
acknowledged.

Instrument Flight Rules (IFR) A set of rules and regulations aircraft and controllers must
operate under when guided by instrumentation (e.g. GPS and Radar). For en
route aircraft this is typically for aircraft above 20,000ft.

National Airspace System Refers to a volume of airspace, and the people, procedures,
equipment, and facilities required to manage it.

NextGen The conceptual next generation air traffic system for the Federal Aviation Ad-
ministration.

Potential Conflict A potential conflict exists between two aircraft when the best avail-
able information predicts that their aircraft trajectories will violate separation
requirements.

Sector A singular unit of airspace under the control of one or two air traffic controllers.

Separation requirement The minimum required distance between any two aircraft. If
separation requirements are violated, then a conflict is deemed to have occured.

SESAR Single European Sky ATM Research (SESAR) is the conceptual next generation
air traffic system for EUROCONTROl.

Taskload The number of tasks or frequency of task occurrence associated with a specific
job description.

Trajectory-Based Operations (TBO) A future operational protocol and concept for man-
aging aircraft. Replacing clearance-based operations, trajectory-based oper-
ations supports longer-term path planning of aircraft. Instead of providing
clearance for individual aircraft actions (heading, altitude, and speed changes),
the air traffic controller approves complete trajectories.

xvii
 CHAPTER I

INTRODUCTION

This thesis is concerned with providing support for the design of decision-support tools to

aid air traffic controllers. The projected growth in air transportation demand is likely to

result in traffic levels that exceed the abilities of the unaided air traffic controller in man-

aging, separating, and providing services to aircraft. From 2010 to 2030, the number of

instrument flight rules (IFR) aircraft operations handled by Federal Aviation Administra-

tion (FAA) en route traffic centers is predicted to increase from approximately 39 million

flights to 64 million flights [35]. Consequently, the FAA, and other air navigation service

providers (ANSP) around the world, are making several efforts to improve the capacity and

throughput of existing airspaces by means of airspace redesign, trajectory based operations,

new traffic flow management tools, and automated data communication and navigation sys-

tems [49, 86]. Ultimately, the stated goal of the Joint Planning and Development Office is

to triple the available capacity of the National Airspace System by 2025 [95].

National Airspace System. In order to address how air traffic controllers can be best

aided, it is important to understand the composition of the National Airspace System

(NAS). The National Airspace System, as operated by the FAA, enables the safe and ef-

ficient transport of aircraft. It consists of a heterogeneous set of interacting units that

includes people, airspaces, technologies, practices, procedures, and policies. Major func-

tional elements of the National Airspace System include the 22 Air Route Traffic Control

Centers (ATRCC), whose responsibilities include managing en route aircraft. A map of the

20 centers spanning the continental United States is shown in Figure 1. Each center is strat-

ified into multiple layers (commonly, low-altitude, high-altitude, and super-high-altitude),

which are further subdivided into en route sectors. Figure 2 provides an example center

map with a subset of the high-altitude sector boundaries for Minneapolis Center (ZMP).

1
 120° W 110° W 100° W 90° W 80° W 70° W

45° N ZSE ZMP

ZLC ZBW
ZAU ZOB
40° N ZDV ZNY
ZOA ZKC ZID
ZDC
35° N ZLA ZAB ZME ZTL
ZFW
30° N ZHU ZJX
ZMA
25° N

Figure 1: Map of the 20 Air Route Traffic Control Centers covering the continental United
States.

Depending on traffic conditions, one or two air traffic controllers are responsible for man-

aging aircraft traffic in a sector. The primary objective of the air traffic controller working

an en route sector is to ensure the proper separation of aircraft at all times. A variety of

approaches and mechanisms exist to improve air traffic management.

Airspace Redesign. In an effort to limit the workload of controllers in the face of in-

creasing traffic demand, airspaces have been redesigned with smaller and more specialized

sectorizations for air traffic controllers to manage. Such sectorization allows air traffic con-

trollers to monitor a smaller set of aircraft, while decreasing variance in traffic patterns.

Examples of such airspace redesign include sectors ZMP15 and ZMP16 in Minneapolis cen-

ter, both which are dominated by aircraft arrivals into Minneapolis - St. Paul International

Airport and are significantly smaller than other sectors in the center, c.f. Figure 2. There

exist limits for which the strategy of sectorization will remain fruitful in limiting controller

workload. The introduction of smaller sectors inherently reduces the amount of traffic that

any single sector can absorb, and places restrictions on the types of conflict-resolution ma-

neuvers available to controllers. Additionally, smaller sectors require additional staffing

and greater coordination between adjacent sectors. The advantage of smaller sectors in

terms of managing controller workload and traffic complexity is reaching its limits. Thus,

2
 100° W 95° W 90° W 85° W

48° N ZMP25
ZMP23 ZMP24
ZMP13
46° N ZMP11
ZMP20 ZMP16 ZMP12
° ZMP19
44 N ZMP15
ZMP18 ZMP17
ZMP29 ZMP30
42° N
ZMP42 ZMP43
°
40 N

Figure 2: Center map for Minneapolis Center (ZMP) with high-alitude sector boundaries.

while airspace redesign, along with the introduction of tighter navigation standards, will

result in further capacity gains, these changes alone are not sufficient to accommodate the

anticipated growth in air traffic demand.

Automated Conflict-Resolution. There has also been a significant investment into the

study and development of aircraft conflict and conflict-resolution systems in order to increase

capacity. Implicitly, conflict-detection algorithms are designed to identify conflicts and en-

able the conflict-resolution programs to generate conflict-free aircraft trajectory solutions.

A number of researchers and ANSPs hypothesize that conflict-detection and resolution tools

will reduce or transform controller workload by decreasing the amount of time and mental

effort controllers spend detecting and resolving potential conflicts [8, 23, 89, 98]. Indeed, the

conflict-detection and resolution process can be time consuming. Under current operations,

for each potential conflict, air traffic controllers are required to discern if intervention is

required; generate resolution commands for one or more aircraft; verbally issue the com-

mands to the aircraft; and monitor and ensure the implementation of the commands. For

potential conflicts between two aircraft, one study estimates the average total time required

3
from controllers, excluding monitoring implementation, is 27.6 seconds [20]. Despite the ap-

pearance of potential reductions in controller workload by displacing some or all separation

tasks to a computer system, there are still concerns that conflict-detection and resolution

algorithms do not aid and could hinder air traffic control performance. Furthermore, prac-

tical implementation has not been achieved. From a human-factors perspective, researchers

have begun to address the topic of ‘allocation of function’ in air traffic control: that is, what

functions and levels of control should be designated to humans or computers. Early results

have exposed a number key benefits and drawbacks of various levels of automation.

Early examples of conflict-resolution algorithms include [28, 69, 70], with a more compre-

hensive survey of proposed models presented in [56]. More recent research and development

in conflict-resolution algorithms seeks to design automated algorithms that provide prov-

ably safe solutions while attempting to satisfy aircraft dynamic limitations [16, 25]. The

majority of these previous developments address a completely automated conflict-detection

and resolution system. Such an automated system is consistent with the proposal, advo-

cated by some air traffic operators and researchers, that the tactical role of radar air traffic

controllers should eventually transition into one of supervisory control of aircraft; an ex-

ample discussion is provided in [33]. In a supervisory control framework, it is the role of

automated tactical conflict-detection and resolution algorithms to ensure separation.

There are significant limitations associated with automated conflict-resolution algo-

rithms. All automated approaches require advanced digital communication and navigation

sub-systems to directly communicate with – and control – aircraft. The development and

implementation of these systems are hindered by the slow uptake of the advanced avionics re-

quired to fully support them. Furthermore, the safety guarantees of these conflict-resolution

algorithms have yet to be established, particularly in accordance with aircraft dynamics,

as all algorithms make use of simplified aircraft models. Such systems are in contrast to

current operations, which are depicted in Figure 3. As it stands, air traffic controllers use

radar systems and vocal communication to manage and issue commands to aircraft. There

are only limited forms of automation, none of which correspond to conflict-resolution. In

these systems, air traffic controllers are responsible for generating resolution commands on

4
 Vocal Communication

Vocal Communication

Radar

Figure 3: Under current operations controllers use radar and vocal communication systems
to track and issue commands to aircraft.

their own.

While there are many proponents for automated conflict-detection and resolution algo-

rithms to automate air traffic systems and separate aircraft, human-factors research has

identified potential hazards. When acting in a supervisory role, under greater levels of

automation, air traffic controllers lose situational awareness and are unable to identify all

future conflicts [64]. Thus, if automated conflict-detection and resolution systems fail to

provide a guaranteed safe solution or the resolution command is not implemented properly,

then it is unlikely an air traffic controller, acting in a supervisory role, will be able to identify

the failure. Furthermore, in the event when air traffic controllers are completely removed

from the system, the lack of fall-back options in the case of failure or nonconforming aircraft

is a prohibitive factor. The option of supervisory control is deemed unsuitable by many

researchers because such a framework ultimately leads to skill atrophy [71, 51]. In cases of

system failure, air traffic controllers lack the proper training and preparedness to respond

appropriately. There are also major concerns about the adverse impact on motivation of

air traffic controllers and the impact on safety.

5
Mixed Automation for Conflict-Resolution. Other efforts have been directed at

mixed-levels of automation in the conflict-resolution process that are more consistent with

human-factors issues. ERASMUS, a EUROCONTROL funded project to study methods

for including automation concepts into air traffic control, introduced the concept of ‘sub-

liminal control.’ In this approach, significant portions of traffic are deconflicted or further

spaced with minor automated speed control commands (±6%); speed-change commands of

this magnitude are not perceptible to pilots or controllers. The use of minor speed changes

to space aircraft reduces both the total number of aircraft that controllers actively mon-

itor for potential conflicts and the number of resolution commands that controllers issue.

There exist numerous publications related to the ERASMUS project and subliminal control

[22, 29, 40]. Early human-in-the-loop simulation studies suggest that controller workload is

reduced.

Despite the potential benefits of subliminal-control, there are concerns about the practi-

cality and ability to implement the systems in the real world: any subliminal-control system

would require advanced data communication and navigation systems to control aircraft di-

rectly. Studies have also indicated that mixed-modes of operation that include dynamic

allocation of separation tasks leads to ambiguity about who is in control and who is respon-

sible for recovery in the case of system failure: pilots or controllers. There is also concern

that the process may lead to loss of job satisfaction for air traffic controllers as a greater

portion of functions is allocated to automation systems [9].

Improving User Displays. Another approach that has received considerable support

from human-factors researchers is the improved design of cognitive and display tools to

provide information in a manner that aids air traffic controllers in parsing information and

coming to quick and appropriate decisions, an example of which is provided in [57]. Overall,

this framework requires little adjustment to the fundamental work practice and methods

employed by controllers. However, it is unknown if the improvement of user displays will

remain suitable at high traffic volumes when taskload demands grows large.

6
 Vocal and Data
Communication

Vocal and Data

Communication

Radar

Decision Support Tool

Figure 4: Conflict-detection and resolution decision-support tools provide possible resolu-
tion solutions to air traffic controllers.

Advisory Decision-Support Tools. The research presented here focuses on another

option touted by human-factors researchers, which is the inclusion of advisory conflict-

detection and resolution decision-support tools to aid air traffic controllers, without re-

placing them. Advisory decision-support tools that provide conflict-resolution options

to air traffic controllers are an alternative to fully- or semi-automated systems that re-

quire advanced data communication and navigation sub-systems. A schematic diagram

of the control structure for human-in-the-loop control with conflict-detection and resolu-

tion decision-support tools is illustrated in Figure 4. In a human-in-the-loop framework,

conflict-detection and resolution decision-support tools identify conflicts and propose reso-

lution commands for the air traffic controller to verify and issue to aircraft. The solution

trajectories can then be uploaded through a data-link into the aircraft flight management

system for approval by pilots.

The major benefits of decision-support tools include an indifference to high traffic

volumes and neutral effects on situational awareness. Much like automated or highly-

computerized solutions, conflict-detection and resolution tools can be designed to handle

7
 Mediating Factors

Quality of
Equipment
Source Factors Result
ATC Complexity:
Air Traffic Individual Controller
Pattern and Sector Differences Workload
Characteristics
Controller
Cognitive
Strategies

Figure 5: Representation of factors affecting controller workload.[66]

large traffic loads, and provide solutions in real-time. Furthermore, because the air traffic

controller is still part of the decision process and has the option to accept or reject pro-

posed solutions, there exists a safety fall-back. However, concerns still remain about the

skill atrophy of controllers and the proper design the decision-support tools. For example,

conflict-detection and resolution systems should be designed appropriately so that human

controllers neither overly depend on the tools nor lack faith in the tool and disregard it

[75, 91].

The inclusion of human-in-the-loop decision-support tools in human-based air traffic

control operations requires a fundamentally different approach to the design and imple-

mentation of conflict-detection and resolution algorithms. The algorithms must explicitly

acknowledge the role of the controllers and accommodate their abilities. Central to deter-

mining the potential success of decision-support tools for separation assurance, research is

required to demonstrate real and meaningful benefits. There is a need to establish that the

inclusion of conflict-resolution decision-support tools will reduce, or at the very least not

worsen, controller workload, while improving service. The research presented here seeks to

understand how the formulation, capabilities, and implementation of conflict-detection and

resolution tools affect the controller. Specifically, the metric of interest is the taskload as-

sociated with the conflict-resolution process, which is implicitly tied to controller workload.

8
Controller Workload. Controller workload associated with air traffic control corre-

sponds to the stress, or “...the amount of effort, both physical and psychological, expended

in response to system demands (task load) and also in accordance with the operators internal

standard of performance” [94].’ The factors leading to controller workload are numerous. In

the review by Mogford et al [66], the authors identify four factors that influence controller

workload: air traffic control complexity, quality of equipment, individual differences, and

controller cognitive strategies. Air traffic control complexity refers to a set of descriptions of

the prevailing air traffic patterns and sector characteristics that aid or hinder in the manage-

ment of aircraft. The next three factors (equipment, individual differences, and strategies)

are referred to as mediating factors. Quality of equipment, for example, considers the accu-

racy of radar systems or the usability of user displays and interfaces for computer-support

tools. Individual differences account for anxiety levels, personality, age, and experience,

which have been demonstrated to correlate with controller workload. And lastly, controller

cognitive strategies incorporate how controllers adjust their strategies in managing traffic

according to pressures from increasing traffic demands. For example, at low traffic volumes

controllers consider a wide variety of information and options in managing traffic, while

at high traffic volumes, controllers limit the amount of information they consider and ad-

just management strategies based on economy, including strategies for resolving potential

aircraft conflicts [92].

Complexity. Early technical practitioners attempted to characterize controller workload

through a set of intrinsic properties associated with source factors, without directly dealing

with the mediating factors. They established the term air traffic control complexity (some-

times simply referred to as complexity) to indicate the relative ease or difficulty of managing

traffic within an airspace. The dynamic density of an airspace is a validated and commonly

used measure of complexity [17, 55, 58, 93]. Initially, the goal of dynamic density was to

match the perceived workload of air traffic controllers by considering a large set of mea-

surable factors (number of aircraft in the airspace, distribution of aircraft speeds, spatial

distribution of aircraft, etc.). Through human-in-the-loop experiments, regression models

9
were established to estimate controller workload. Kopardekar et al. [55] present a review

of regression parameters used to compute the dynamic density. While only accounting for

half the variance in controller workload, dynamic density established that the number of

potential conflicts and aircraft maneuvers, in addition to the local density of aircraft and

other measures, are strong factors in predicting controller workload.

More recent efforts seek to establish intrinsic measures of complexity using different fea-

tures than those accounted for in dynamic density. Prandini et al. [78] propose probabilistic

air traffic complexity measures based on aircraft occupancy within the airspace. A similar

probability model based on traffic flows is proposed in [84]. The authors demonstrate how

parameterized generative models can be used to create airspace maps to highlight regions

of high traffic or regions where aircraft interactions and conflicts are more likely. Delahaye

and Puechmorel [24] propose a complexity measure based on non-linear dynamical systems.

Their measure of Lyapunov Exponents captures the structure in a situation by identifying

the organization of trajectories in traffic patterns.

A common theme to the previous works is that they all implicitly consider aircraft

conflicts and interactions. Dynamic density acknowledges aircraft interactions by including

the spatial density of aircraft and number of potential conflicts in its measure. Furthermore,

approximations of the conflict-resolution process is partially accounted for by including

aircraft altitude, heading, and speed changes. These maneuver actions are often in response

to potential conflicts or a desire by controllers to further space aircraft to ensure clear

separation. The second set of complexity measures discussed are based on either a measure

of interaction (i.e. probability of conflict) or measure of organization of the traffic flow.

Regardless, all the previous methods implicitly consider potential conflicts, and the resulting

conflict-resolution process.

There should be no surprise that aircraft interactions and conflicts are sources of com-

plexity and controller workload. The primary objective of the air traffic controller is to

ensure the proper separation of aircraft, thereby avoiding collisions, at all times. Thus,

conflict resolution takes precedence over any other management tasks.

10
Research Needs. In an effort to understand the potential workload implications of in-

troducing conflict-detection and resolution tools into air traffic control systems, a detailed

study of the conflict event process and the implementation of the separation algorithms

are necessary. Specifically, the research presented here addresses the conflict taskload re-

quirements to resolve traffic. That is, the studies contained in this thesis examine how

many resolution commands air traffic controllers issue in order to separate traffic under the

guidance of a conflict-detection and resolution decision-support tool. Such an approach is

valid given the importance of the conflict-resolution process to influencing controller work-

load, as implicitly validated by dynamic density and other methods. Addressing taskload

requirements is something previous researchers have considered in limited forms, even in

regards to the development of conflict-resolution algorithms.

Studying and modeling the conflict-detection and resolution event process enables an as-

sessment of the number of potential conflicts air traffic controllers can expect to encounter,

the structure of the potential conflicts (i.e. how many aircraft are involved), and the manner

in which the potential conflicts occur in time. Furthermore, analysis of potential conflicts fa-

cilitates a discussion of the best practices for designing and implementing conflict-detection

and resolution systems in an effort to regulate the number of resolution commands used to

space aircraft.

The method by which conflict-detection and resolution algorithms identify and resolve

conflicts is critical to the function of a decision-support tool. Much like human air traffic

controllers assess potential conflicts and determine conflict-resolution solutions, any human-

in-the-loop decision-support tool will be required to do the same. The process by which these

tasks occur is not straightforward. The ability to identify potential conflicts depends on the

accuracy of trajectory prediction and aircraft position reporting systems. Meanwhile, the

guarantee that any resolution command ensures conflict-free travel for aircraft depends on

the accuracy of the conflict-detection algorithms and the strength of the conflict-resolution

algorithms. Previous development in conflict-resolution algorithms has largely overlooked

the accuracy and availability of information in regards to trajectory prediction and con-

flict detection. Ultimately, any conflict-detection and resolution algorithm must consider if,

11
when, and what resolution commands should be issued. These considerations are a func-

tion of both the information available and the policy under which the conflict-resolution

algorithm operates.

1.1 Thesis Objectives

In the face of increasing air traffic, there is a growing need to design systems to support,

not replace, air traffic controllers in managing and separating aircraft. If human-in-the-

loop decision-support tools for conflict-detection and resolution are to be integrated into

the repertoire of air traffic controllers, then it is vital that the tools operate in a manner

consistent with human work practice. Thus, at a minimum, designers of conflict-detection

and resolution algorithms must account for taskload implications inherent in any design.

The objective of this thesis is to explore and accomplish the following:

• Generate a method to model traffic in order to assess the conflict event-process within

an airspace.

• Abstract the behaviors and characteristics of conflict-detection and resolution systems.

According to the abstractions, analyze how their implementation and capabilities

affect controller taskload.

• When possible, establish best-practices for designing conflict-detection and resolution

algorithms to manage the number of resolution commands used to prevent air traffic

conflicts.

By researching the topics above, the major contribution of the thesis is to derive per-

formance reference models that serve as a comparison tool, and ideally, to establish best

practices for the design of real-world conflict-detection and resolution systems. Lastly, un-

derstanding how the design of conflict-detection and resolution systems affects controller

taskload provides a basis to improve models for predicting controller workload.

1.2 Thesis Outline

The thesis is divided into two major sections. The first major section, begins with Chapter

2 introducing a procedure for modeling uncontrolled air traffic within an airspace. The

12
resulting traffic model serves as a basis for the remainder of the thesis, except where stated.

The traffic model enables the analysis of the conflict event process presented in Chapter

3, where the temporal and spatial distributions of conflict events are explored. Multiple

airspaces are studied to demonstrate that the design and structure of an airspace results

in different levels of air traffic control complexity. The analysis in Chapter 3 motivates the

need for advisory conflict-detection and resolution decision-support tools.

Following the initial analysis, the second major section, beginning with Chapter 4,

proposes a graph-based methodology for modeling the conflict-detection and resolution

process. Next, in Chapter 5, a framework for abstracting conflict-detection and resolution is

described. The abstraction considers the implementations and design features for conflict-

detection and resolution algorithms. Using the uncontrolled traffic models developed in

Chapter 2 and the abstraction of conflict-detection and resolution systems in Chapter 5,

a process for executing traffic simulation is described in Chapter 6. From the simulations

of the graph-based conflict-detection and resolution algorithms, taskload requirements for

air traffic controllers using these tools are analyzed and best practices are proposed in

Chapter 7. Finally, a summary and conclusions on the design of conflict-detection and

conflict-resolution algorithms are discussed in Chapter 7.

To aid in navigation of the thesis, a representation of the chapter dependencies is shown

in Figure 6.

13
 Chapter 1
Introduction

Chapter 4
Chapter 2
Graph Model
Traffic Model of Traffic

Chapter 3 Chapter 5
Initial Conflict Abstraction of
Analysis Tools

Chapter 6
Simulation
Implementation

Chapter 7
Simulation
Analysis

Chapter 8
Conclusion

Figure 6: Chapter dependency

14
 CHAPTER II

AIRSPACE MODEL

In this chapter, a methodology for generating an uncontrolled open-loop air traffic model

is presented. The resulting uncontrolled traffic model approximates air traffic within an

airspace as if aircraft flew irrespective of each other and without controller input. The

uncontrolled air traffic model enables characterization of the conflict-event process within

a sector and is later used to assess the performance of conflict-detection and resolution

algorithms. Specifically, the uncontrolled air traffic model attempts to replicate potential

conflicts that would be commonly resolved in real-world operations - in both location and

aircraft configuration and structure. The traffic model is based on a re-sampling of historical

aircraft radar data.

Using resampled historical radar data, scenarios of mock 4D uncontrolled flight trajec-

tories are created that include variations over a range of variables. The goal is to replicate

variance that is commonly found in air traffic, even for aircraft flying along the same path.

In practice, variations in aircraft trajectories commonly appear in aircraft climb rates and

speeds. These differences are due to airline and pilot preferences, as well as present atmo-

spheric conditions, aircraft and engine capabilities, flight management systems (FMS), and

aircraft weight.

The methodology presented here to model uncontrolled air traffic is an alternative to

collecting historical flight plan information and simulating aircraft traffic along the flight

plans with stochastic properties, similar to that in [3, 10]. It is also an alternative to using

the software described in [10], which was inaccessible due to security restrictions.

The methodology presented here to model uncontrolled air traffic is an alternative to

collecting historical flight plan information and simulating aircraft traffic along the flight

plans with stochastic properties, similar to that in [3, 10]. By re-sampling aircraft trajecto-

ries, proprietary aircraft and engine manufacturer data, as well as sophisticated simulation

15
software are not required. The methodology described here is an alternative to using the

software described in [10], which was inaccessible due to security restrictions. Properties of

the methodology for generating high-intensity traffic are verified to be consistent with the

software in [10].

The traffic model created in this chapter is used throughout the remainder of the thesis,

except where noted. First, in Section 2.1 a list of assumptions required for the development

and the use of the model is given. In Section 2.2, a description of the data set used to

construct the uncontrolled open-loop model is provided. Next, Section 2.3 details the pro-

cedure for creating sample traffic days of uncontrolled air traffic. Section 2.4 defines how

potential aircraft conflicts are identified. Finally, verification of key properties is provided

in Section 2.5.

2.1 Assumptions

The open-loop traffic model generated in this thesis attempts to represent uncontrolled air

traffic within an airspace, as if aircraft flew irrespective of each other, and without controller

input. The traffic model is based on a re-sampling procedure of historical flight data. The

traffic model is based on the following assumptions:

• A traffic pattern can be parameterized by the origin and destination of each flight. Re-

sampling flights to maintain the same origin-destination pairs results in traffic cases

that mimic the original spatial traffic pattern. Replicating the spatial traffic pattern

preserves the overall experience of the air traffic controller.

• Scaling traffic by reducing the time frame over which it occurs is equivalent to in-

creasing the total amount of traffic (i.e. simulating 24 hours of traffic over 8 hours is

similar to three times the traffic in a 24 hour time period). By extension, a 20 minute

time period with triple the traffic demand can be scaled over a longer time period to

compute taskload demands over a complete 24 hour day.

• Origin-destination demand pattern, aircraft models, sector boundaries, and routes

remain similar as traffic intensity increases. Or implicitly, in 20 years when traffic

16
 volumes have increased, the traffic, aircraft, and sector properties remain similar.

From a practical perspective, this is a tenuous assumption, as almost surely, fleet mixes

and sector boundaries will change in time. Furthermore, there is limited understanding

of the impact of unmanned aerial vehicles on air traffic management.

• The sampling procedure models uncontrolled air traffic in a manner that allows for

valid assessment of the conflict-event process.

• The validity of the uncontrolled traffic model in an airspace is preserved despite pos-

sible controller intervention of aircraft in adjacent airspaces.

2.2 Data Set

The traffic model of uncontrolled flight trajectories is based on 28 days of Performance Data

Analysis and Reporting System (PDARS) data. The PDARS data set includes aircraft radar

position information (latitude, longitude, altitude) nominally sampled every 12 seconds,

covering the dates May, 21, 2007 through June 17th, 2007, over Minneapolis Center (ZMP).

Each day of flight data is compiled separately. The complete data set contains 107,671

high-altitude flights that pass at or above 20,000 ft at some point in their trajectory; this

includes both general aviation and scheduled flights. A single day of traffic includes data

for flights departing their origin on the stated day. Thus, a flight leaving from California

late-night may appear inside ZMP within the 25th hour of the day.

A presence map of the traffic over the the center is presented in Figure 7 for June 14th ,

2007. Darker regions of the map indicate areas of high traffic; conversely, low traffic regions

are lighter in color. Over the 28 days, variations in the traffic pattern exist due to weather

and wind conditions, as well as origin-destination demand. However, qualitatively, the daily

traffic pattern remains similar over the complete data set.

Minneapolis center is divided into 6 major areas according to function, traffic type,

and volume. The area boundaries are illustrated in Figure 8. Each area is further sub-

divided into sectors; each sector spans a given area and altitude bound. Typically, the

busier areas, by aircraft count, are located in the southern regions of the center, while the

least busy areas are in the north. Accordingly, the sector divisions within each area are

17
Figure 7: High-altitude traffic pattern for June 14th , 2007 in Minneapolis Center.

102° W 99° W 96° W 93° W 90° W 87° W 84° W

48° N
Area 4
Area 1
45° N Area 2
Area 3
Area 5
42° N
Area 6

Figure 8: High-altitude boundaries for Areas in ZMP.

18
 98° W 96° W 94° W
42° N

ZMP43
41° N ZMP42 ZMP40 ZMP39
ZMP38

40° N ZMP38, ZMP39

ZMP42, ZMP43
ZMP40

Figure 9: High-altitude sectors boundaries in Area 6 of ZMP.

different. For some sectors in Area 1 and Area 4, where traffic volumes tend to be light,

an air traffic controller manages the airspace from the ground up, excluding any special

airspaces (restricted, military, airport, etc.). In contrast, Area 6, contains highly stratified

sectors; the high-altitude sector boundaries are illustrated in Figure 9. Two low altitude

sectors cover the airspace from ground-level to just under FL240. High altitude airspace is

divided by Sectors 38 and 39 controlling aircraft from FL240 to FL340, and Sectors 42 and

43 controlling aircraft from FL340 to FL390. Finally, the air traffic controller for Sector 40

is responsible for managing all aircraft at FL400 and above in Area 6. The vertical division

of Area 6 is an example of the sectorization of en route centers that is reaching its useful

limits.

Traffic within the center fluctuates over the course of a day and from day to day. The

25th , 50th , and 75th percentiles for the daily aircraft traffic counts in the center are given

in Figure 10. While keeping a relatively consistent shape, the demand does vary between

days. Additionally, demand is seasonally varying, especially in the northern areas [36].

Accordingly, while the air traffic model presented in this chapter seeks to aid in modeling

the conflict-detection and resolution process, it must be understood that the resulting model

cannot exactly represent all days, as variation between days and within days exists.

19
 Aircraft Count in Center
150

100
25th Percentile
50 50th Percentile
75th Percentile
0
0 5 10 15 20 25
Hour of Day
Figure 10: Diurnal aircraft counts within Minneapolis Center for the dates May, 21, 2007
through June 17th, 2007.

2.3 Sampling Procedure

To generate potential conflicts based on realistic air traffic patterns, flight data from a seed

day is used to create 100 new scenarios of uncontrolled flight trajectories. A scenario of

uncontrolled flight data approximates the traffic pattern of the seed day. The uncontrolled

flight trajectories represent flight paths that aircraft would fly irrespective of other aircraft,

and as if air traffic controllers provided no input. In generating the 100 traffic scenarios,

flights from all days, excluding the seed day, are sampled according to the origin-destination

pairs of the seed day to populate the flight trajectories. By sampling according to origin-

destination pairs found in the seed day, the overall traffic pattern within the center is

preserved. Preservation of other key metrics is demonstrated in Section 2.5. Based on the

100 traffic scenarios, the timespan over which the flights occur is reduced to increase the

traffic intensity in the center.

The first step in the resampling process is to index all flights according to origin-

destination pairs, and to identify the entry and exit times of the aircraft. In the PDARS

data, the origin and destination information of the flight plan is described according to fixes

(e.g. NASAL), airports (e.g. KATL), and latitude-longitude data (e.g. 4510N/10354W).

Origin and destination pairs that are described by fixes and airports are indexed directly.

For approximately 4% of the flights (4329 trajectories), the origin data is described by

latitude and longitude. In these cases, the origins are manually clustered by location and

20
 104° W 100° W 96° W 92° W 88° W 84° W 80° W
50° N

48° N

46° N

44° N

42° N
Latitude / Longitude Origin
40° N
Cluster
38° N Center Boundary

Figure 11: Spatial clustering of origins described by latitude and longitude.

indexed sequentially. Any latitude-longitude origins that are not clustered form an ‘outlier’

cluster. Figure 11 illustrates the complete set of origins described by latitude and longitude,

and the manual cluster partitions.

Following clustering and indexing, a total of 11,014 origin-destination pairs are iden-

tified. Ranking the origin-destination pairs according to the number of occurrences, the

frequency distribution, as well as the probability and cumulative distributions, are depicted

in Figure 12. As an example, the most common origin-destination pair, ROBBY to KMSP,

occurs 2442 times. The fix ROBBY is located outside and to the south of ZMP, near the bor-

der with Chicago Center; KMSP corresponds to Minneapolis St. Paul International Airport.

According to the relationships illustrated in Figure 13, 224 origin-destination pairs repeat

at least 100 times over 28 days and account for ∼ 52% of the traffic. Origin-destination

pairs that occur less than 10 times are declared to be rare. Rare origin-destination pairs

account for 15% of all flights, yet constitute 9639 of the 11014 pairs. The classification

of rare origin-destination pairs is later used by the sampling procedure in creating new

uncontrolled flight trajectories.

For a seed day with M flights passing through the center, the origin-destination pairs
s = {(O s , D s ), . . . , (O s , D s )}, and the arrival
are given by the set of ordered pairs, Pod 1 1 M M

21
 1
3 −2
10 10

Cumulative Distribution
Probability Distribution
0.8

2
10 0.6
Count

0.4
1 −4
10 10
0.2

0
10 0
0 5000 10000 0 5000 10000 0 5000 10000
Origin−Destination Pair Origin−Destination Pair Origin−Destination Pair
Figure 12: Distribution for origin-destination pairs for flights within Minneapolis Center.

100
4200
Number of OD Pairs

1535 75 Percent of flights

561
205
75 50
27
10 # of Origin−Destination pairs 25
Percent of flights
0 1 2
03
10 10 10 10
Number Origin−Destination pair repetitions

Figure 13: Clustering relationships between repeated origin-destination pairs and aircraft
fights.

22
times into the center T s = {T1s , . . . , TM
s } for all flights. Here, the superscript ‘s’ indicates

that the variable relates to the seed day. For an arbitrary ith day, the origin-destination
i and T i are defined in a similar manner. Furthermore, all flights
pairs and arrival times Pod

are indexed for each day, such that Fji = {(x1 , y1 , z1 , t1 ), . . . , (xk , yk , zk , tk )} corresponds to

the space-time trajectory points for the j th flight of the ith day. The spatial values are

given in longitude (degrees), latitude (degrees), and altitude (x100ft); time is measured in

seconds from midnight (Coordinated Universal Time) on 1/1/1970, i.e. unix time. For

each new traffic scenario, the re-sampling procedure assigns origin-destination pairs and

arrival times to P rs = {(O1rs , D1rs ), . . . , (OM

rs , D rs )} and T rs = {T rs , . . . , T rs }, along with
M 1 M

flight trajectories to F rs , thus, generating a resampled traffic model. The procedure for

populating the resampled traffic scenarios is now detailed.

When resampling the seed day with uncontrolled flight trajectories to create each traffic

scenario, the first step is to set P rs = P s , thereby preserving the traffic demand pattern

within the center. Next, arrival times for each of the resampled trajectories are assigned

according to the original arrival times of the seed aircraft, and the desired traffic intensity.

For the resampled model, the arrival time for each aircraft includes a normal random term,

σN , and is inversely scaled by I to increase the traffic intensity. That is,

Tjrs = (Tjs + σN )/I, (1)

By adding the random normal variable σN to the original arrival times, Tjs , arrival times

are slightly perturbed. When I = 1, the scaling value implies that the traffic intensity

of the resampled traffic case matches the seed day. For I = 3, the traffic case is three

times the traffic intensity. Dividing by the intensity term I compresses the time frame over

which traffic is simulated. Later in the thesis, analysis focuses on the number of potential

conflicts in a traffic scenario. A multiplicative factor is required to scale conflict counts for

the compressed traffic cases (i.e. I > 1). Scaling allows study of the compressed traffic

cases over a standard 24 hour day. Otherwise, unequal day lengths prevent meaningful

comparison between any two different traffic intensities. For the traffic scenarios generated

in this thesis, σ = 5 minutes.

23
 For each flight trajectory Fjrs in F rs , a new flight trajectory is sampled from all other

flights with the same origin-destination pair given by (Ojrs , Djrs ). If an arbitrary origin-

destination pair, (Ojrs , Djrs ), is classified as rare, then the origin-destination pair and the

new trajectory is resampled from all origin-destination pairs that are also rare and that occur

within 15 minutes of the original arrival time of the seed aircraft, Tjs . By constraining rare

flights to be sampled within a 30 minute window of the original arrival time, the approximate

behavior of traffic around the same time of day is preserved. The time values for the space-

time trajectory are reset, starting with the resampled arrival time Tjrs . Thus, if the sampled

flight originally contained trajectory points for times {t1 , . . . , tk }, then the new times for

the resampled flight trajectories are {Tjrs , t2 − t1 + Tjrs , . . . , tk − t1 + Tjrs }. Through this

sampling procedure, all flight trajectories are decorrelated and do not contain controller

actions relative to one another. Each traffic scenario can then contain potential conflicts.

The last steps of the sampling procedure include parsing aircraft trajectories according

to the sectors they pass through. First, each flight trajectory is interpolated at 1 second

increments. Portions of a trajectory that pass through a given sector are isolated. The

isolated trajectory is then projected into a local coordination frame using the standard

gnomonic projection about the centroid of the sector. The projection generates a local

coordination frame in units of NM for the X-Y axis. Also calculated are the arrival and exit

times of the aircraft for the sector. The process is repeated for each sector and for all flights

in the resampled traffic scenario. The final result is a set of trajectories that corresponds

to a traffic scenario for a specified sector. Traffic scenarios for individual sectors are of

particular interest because each airspace corresponds to a control volume managed by a

single air traffic controller.

The complete resampling procedure provides a method for generating uncontrolled flight

trajectories through each sector. Initialized with a large set of aircraft trajectories and

origin destination data - selecting a seed day, traffic intensity and sector - the output of the

resampling procedure is 100 traffic scenarios with aircraft trajectories passing through the

specified sector.

24
2.4 Separation Detection

For the purposes of determining controller-action taskloads associated with the conflict-

resolution process, relative aircraft distances are checked to determine if aircraft pairs come

in close proximity. According to the relative proximity of aircraft pairs, action by the

controller may be required. In this section, a mathematical description of the time and

distance at which aircraft come in close proximity is provided. The proximity distances and

times are calculated for all traffic scenarios and traffic intensities. Later chapters of this

thesis make use the distances and times as estimated measurements for detecting potential

conflicts.

Given a sector and traffic scenario, aircraft separation distances are calculated. Denote

the trajectory of an arbitrary aircraft Ai in the local frame by the parametric functions

(xi (t), yi (t), zi (t)), representing the position and altitude of the aircraft in time. For two

aircraft, Ai and Aj , the separation distance Di,j (t), at any point in time, is

p
 (xi (t) − xj (t))2 + (yi (t) − yj (t))2 if |zi (t) − zj (t)| < 1000


Di,j (t) = (2)

∞

else.

According to standard separation requirements for the en route environment, a conflict

between two aircraft Ai and Aj is said to occur if the aircraft come within 5NM laterally,

and 1000 ft vertically of each other, i.e. when Di,j (t) < 5NM for some value of t. Thus,

if a conflict-detection probe indicates the potential for a conflict in the near-term (e.g. 5

minutes), then the air traffic controller should issue resolution commands to at least one

aircraft to ensure a conflict does not occur. Otherwise, without resolution, conflict is likely.

Often times however, a more conservatively safe approach is taken in determining if a

resolution maneuver is required. How a conflict-detection tool classifies a potential con-

flict between two aircraft depends on the certainty of trajectory prediction tools and their
miss , is
uncontrolled minimum miss-distance. The uncontrolled minimum-miss distance, Di,j

defined as

miss
Di,j = min Di,j (t) (3)
t

25
 Di,j>15 NM
320
Di,j<15 NM
Flight Level 310
Di,j<12 NM
300 Di,j<10 NM
290 Di,j<8 NM
Di,j<5 NM
280
−60

−40
−80 −100
−20 −40 −60
NM NM
Figure 14: Example conflict between two aircraft.

prior to any action by the air traffic controller.

An example of an uncontrolled conflict is provided in Figure 14. The figure illustrates

two aircraft breaking the 5NM mile separation requirement. In this case, one aircraft is

ascending, while the other aircraft is in level flight. In the figure, as the two aircraft close

in lateral and vertical distance, the trajectories lines grow thicker and become darker and

more red in color. After passing each other, the thinner green segments of the trajectories

indicate that the aircraft are a safe distance apart.

For the purposes of this thesis, the time of the uncontrolled minimum miss-distance is

of interest. For the aircraft pair Ai and Aj , the critical time is denoted by tci,j , and given

by

tci,j = argmin Di,j (t). (4)

t

The critical times values are later used in Chapter 6 to indicate the predicted time of

potential conflict between two aircraft when simulating conflict-detection and resolution

algorithms.
miss , and
Based on the mathematical descriptions above, the minimum miss-distances, Di,j

the critical times, tci,j , are calculated for all pairs of aircraft. For aircraft pairs with minimum

miss-distances less than 15NM, qualitative and quantitative information pertaining to the

26
flights is saved (e.g. flight-phase, conflict-angles, groundspeed). The information is later

used by conflict-detection and resolution models to determine to which aircraft maneuvers

should be issued commands to resolve potential conflicts.

2.5 Application and Verification of the Sampling Procedure

The resampling procedure is applied using June 14, 2007 as the seed day. Over the 28 days

of traffic, June 14th corresponds to the day with the largest observed traffic load. Thus, the

day provides a nominal lower bound on the maximum capacity of the center. In line with

the goal of the Joint Office of Planning and Development, which envisions a tripling of the

National Airspace System (NAS) capacity [49], increasing the traffic intensity for June 14th ,

according to the resample procedure, provides a reasonable test-bed. For the remainder of

the thesis the traffic scenarios of uncontrolled aircraft trajectories generated from the seed

day are used for study. In this section, the center-wide traffic scenarios are tested to support

the assumption that the resampling procedure yields traffic scenarios similar to the seed

day. In particular, the spatial, aircraft model, and maximum altitude distributions are

compared against the seed day using statistical methods.

Because the statistical tests are only concerned with the spatial, aircraft model, and

altitude distributions, they only need to be applied to the traffic scenarios for a single

traffic intensity. Each of the metrics is independent of the time scaling found in Equation 1,

and as such, any statistical results for one traffic intensity holds for all traffic intensities.

2.5.1 Spatial Distribution of Aircraft

To verify that the 2D spatial distribution of the resampled uncontrolled trajectories match

that of the seed day, the distributions are compared using a two-sample Kolmogorov-

Smirnov test (KS test). The two-sample KS test is used to verify if two empirical dis-

tributions are sampled from the same distribution. The null hypothesis and alternative

hypothesis are as follows:

27
Figure 15: Spatial distribution of aircraft for June 14th , 2007 and a generated traffic scenario.

Ho : Samples come from the same distribution.

Ha = : Samples come from the different distributions.

Rejection of the null hypothesis is quite meaningful, as it implies that the resampled

traffic scenarios are not representative of the seed day. Instead of the more stringent α = .01

value, setting α = .05 takes a more conservative approach in determining if traffic scenarios

are similar.

The two-sample KS test requires single-dimensional empirical distributions. As such,

spatial aircraft counts over the center must be reshaped to be consistent with testing. The

single-dimension empirical distribution for the seed day is generated by first overlaying a

set of 5NM by 5NM bins over the airspace. For each aircraft that passes over an arbitrary

bin bi , the single-dimension sampled distribution includes the value i. For example, if 64

aircraft pass through the piece of airspace corresponding to bin, b13 , then the value 13 is

included in the sampled empirical distribution 64 times.

For the historical June 14th seed day, the empirical spatial distribution of aircraft within

28
the center is represented by the frequency map in Figure 15. Each bin index i is calculated

according to the X and Y indicies generated by the 5NM by 5NM grid. As an example,

the figure also includes the spatial distribution of aircraft for one of the traffic scenarios

generated using the resampling procedure. Applying the KS test (with α = .05) between

the seed day and the 100 traffic scenarios, the null hypothesis is never rejected. Thus, it

cannot be stated with certainty that the sampled spatial distributions come from different

populations.

The KS test supports the assumption that the spatial traffic patterns of the resampled

traffic scenarios are representative of the original seed day (see Section 2.1). However,

despite accepting the null hypothesis for each case, any spatial correlation in the empirical

distributions cannot be accounted for in the verification procedure due to reordering of the

bins. Additionally, any correlations between aircraft flight trajectories cannot be verified

by the KS test.

2.5.2 Maximum Altitude Distribution of Aircraft

The previous verification test checked that the spatial distribution of aircraft trajectories

in the X-Y plane for the seed day and generated traffic scenarios are similar. To complete

verification of the spatial distribution of aircraft, another study is required that considers

the vertical direction. Division of the test into the planar and vertical directions is re-

quired because a complete comparison of the 3D spatial distributions is not meaningful; the

sparseness of the trajectories ensures that the possibility of rejecting the null hypothesis is

extremely unlikely. To consider the vertical distribution of aircraft, the maximum attained

altitudes of the seed day and generated traffic scenarios are compared.

For each flight, the maximum altitude attained by each aircraft is extracted. Let

Asmax = {As1 , . . . , AsM } be the set of maximum altitudes for each aircraft in the seed day,

where the maximum altitude of the j th flight is given by Asj = max(zi | (xi , yi , zi , ti ) ∈ Fjs ).

Similarly, let Ars

max be defined for each of the resampled traffic scenarios. The distribution

Asmax for the seed day and Ars

max for one of the generated traffic scenarios are illustrated in

Figure 16. As shown in the figure, the distributions of maximum altitudes between the two

29
 600
Historical

Frequency
400 Generated

200

0
200 250 300 350 400 450 500
Maximum Altitude [x100ft]
Figure 16: Distribution of maximum aircraft altitudes for June 14th , 2007 and a generated
traffic scenario.

data-sets are similar. When applying the KS test to verify or refute if the samples come

from the same distribution, the null hypothesis is accepted. More so, the null hypothesis is

accepted for the 100 generated scenarios.

2.5.3 Distribution of Aircraft Models

Because the resampling procedure is based on origin-destination pairs, there is no guarantee

that the aircraft model types in the generated traffic scenarios are representative of the seed

day. Thus, the two sample KS test is applied to verify or refute that the distribution of

aircraft models are similar. Maintaining similar distributions of aircraft models is important

because it is the variation in aircraft performance (often dependent on aircraft model)

that results in differences in aircraft dynamics (especially climb rates). Furthermore, the

decision-making process for conflict-resolution (for both automated systems and human air

traffic controllers) must consider the dynamic capabilities of aircraft in implementing any

resolution commands.

The distribution of the top 20 aircraft models found in the seed day are shown in

Figure 17. The figure also includes the frequency of occurrence for the same 20 aircraft

in one of the generated traffic scenarios. When applying the two-sample KS-test to the

complete distribution of aircraft models, the null hypothesis is accepted; the empirical

distribution of aircraft model types for the seed day and the generated traffic scenarios

are similar. Matching aircraft model distributions is an example of how the resampling

30
 500
400 Historical

Frequency
300 Generated
200
100
0

A320
A319

MD83
E145
MD82
B752

CRJ2
B737
B738
B733
CRJ7
DC93
B763
B744
B772

H25B
C750
B722
C560
DC95
Aircraft Model

Figure 17: Distribution of aircraft models for June 14th , 2007 and a generated traffic sce-
nario.

procedure maintains specific properties despite the fact that the procedure only considers

origin-destination pairs in creating new traffic scenarios.

2.6 Scaling for Equitable Comparison

As the traffic intensity increases, i.e. for I > 1, the length of each simulation day decreases

inversely proportional to the intensity factor. A result of the compressed time scale is that

aircraft are more closely spaced, and thus, the likelihood for potential conflicts increases.

Problematic however, is that when considering the number of potential conflicts in each

traffic scenario across traffic intensities, unequal lengths of day prevent fair comparison

directly; the number of conflicts in an 8 hour day (3X intensity) is not equivalent to the

number of conflicts in a 12 hour day (2X intensity) or 24 hours day (1X intensity). In

this section, a scaling factor applied to the compressed days is proposed, thereby allow-

ing equal comparison of conflict counts across all traffic intensities. Validity of a scaling

factor is tested through a process that extrapolates conflict totals from subsampled traffic

data. Additionally, in the process of verifying the scaling factor, the expected number of

uncontrolled conflicts is demonstrated to increase quadratically with traffic intensity.

The primary metric of interest in this section is the mean total number of uncontrolled

conflicts that occurs in a 24 hour day for each traffic intensity level within a sector. The

labeling of an uncontrolled conflict follows the definition provided in Section 2.4; a conflict
miss (t), is less than 5NM
occurs when the minimum-miss distance between two aircraft, Di,j

31
at some point in time. Uncontrolled conflict totals at the sector level, not the center level,

are of particular interest; controllers are responsible for managing and separating aircraft

in a single sector. While the mean conflict totals for compressed traffic intensities I > 1 can

be extracted for the native shortened simulation day, a scaling factor is required to estimate

the total number of conflicts over a 24 hour day. Accordingly, the following hypothesis is

made:

Hypothesis 1. Denoting N̄CI,C to be the average number of uncontrolled conflicts for the

traffic intensity I > 1 for the compressed day, and E[NCI,24 ] to be the expected number of

conflicts for the traffic intensity I > 1 over a 24 hour period, then

E[NCI,24 ] = N̄CI,C × I. (5)

Further, it is hypothesized that E[NCI,24 ] can be modeled by the equation

E[NCI,24 ] = k1 I 2 + k2 I. (6)

To support the hypothesis above, a subsampling procedure is used to create traffic

scenarios with traffic intensities I < 1. The procedure for generating subsampled traffic

scenarios maintains a 24 hour day, thereby allowing equal comparison of conflict-counts

between other subsampled intensities. Next, the average conflict totals of the subsampled

traffic are extrapolated for higher intensity values (i.e. I > 1) to demonstrate that the

expected number of conflicts in a 24 hour day for the compressed traffic scenarios matches

the hypothesized values of E[NCI,24 ].

The process for generating the subsampled traffic is detailed in Procedure 1, and de-

picted in Figure 18. In the procedure, the variables P rs , F rs , T rs correspond to the origin-

destination pairs, trajectories, and aircraft arrival times of a resampled traffic scenario with

traffic intensity I = 1. These variables were previously generated using the process outlined

in Section 2.3. For the subsampled traffic intensity Is < 1, a new traffic scenario is cre-

ated by randomly removing aircraft and trajectories from the original 1X resampled traffic

scenario. Figure 18 is an example depiction of how the number of uncontrolled conflicts is

reduced as aircraft are removed from a traffic scenario. The aircraft that remain are used to

32
 A4 A4 A4

A5 A5

A1 A1 A1
A2 A2 A2
A3 A3

Flight plan Removed aircraft

Uncontrolled conflict Affected conflict Affected conflict
(a) Original scenario (b) Aircraft subsampling (c) Resulting scenario

Figure 18: Illustrative example of the subsampling procedure.

populate the variables P rs,Is , F rs,Is , T rs,Is , which corresponding to the origin-destination

pairs, trajectories, and aircraft arrival times of the subsampled traffic scenario.

Procedure 1 Subsampling Procedure

N = card(P rs )

c1 = 0, c2 = 0

while c1 <= N do

c1 = c1 + 1

if rand < Is then

c2 = c2 + 1

Pcrs,I
2
s
= Pcrs
1

Fcrs,I
2
s
= Fcrs
1

Tcrs,I
2
s
= Tcrs
1

end if

end while

Application of the subsampling procedure to the 1X traffic scenarios results in subsam-

pled traffic scenarios that occur over a 24 hour time period and contain a fraction of the

33
 400 Interpolated from
subsampled data

Total Number of Conflicts

Subsampled
300
Compressed

over 24 Hrs
200 ZMP42 ZMP28

100 ZMP22

0
0.5 1 1.5 2 2.5 3
Traffic Intensity
Figure 19: Extrapolated conflict totals based on sub-sampled traffic scenarios.

original flights. For a subsampled traffic intensity value of Is = .5, the expected number of

aircraft in each subsampled traffic scenario is half that in the original resampled 1X traffic

scenario. Again, because the uncontrolled conflict totals in each case corresponds to a 24

hour time period, no scaling is required for direct comparison of uncontrolled conflict totals

between intensities.

For traffic intensities I = [.3, .4, .5, . . . , 1], the average total number of uncontrolled

conflicts occurring over the 24 hour simulation is illustrated in Figure 19 for sectors ZMP22,

ZMP28, and ZMP42. The resampling process for creating the 1X traffic scenarios, and the

subsampling procedure for generating lower traffic intensity scenarios, yield traffic scenarios

occurring over 24 hour time periods.

A least-squares fit is applied to the 1X and subsampled average uncontrolled conflict

totals for the 24 hour day, assuming a quadratic model of the form

E[NCI,24 ] = k1 I 2 + k2 I + k3 , (7)

where I corresponds to the traffic intensity. At I = 0, the traffic intensity implies that there

are no aircraft in the sector and hence no potential conflicts. Accordingly, the y-intercept

value for the quadratic model is set to k3 = 0. The remaining coefficient terms, k1 and k2 ,

34
are calculated by the least-squares solution

K = (AT A)−1 AT b (8)

where,    
2 .3,24
.3 .3 E[NC ]
   
E[NC.4,24 ]
 2   
.4 .4  
A= , b =   , K = k1 k2 .
   
 .. ..   .. 
 . .  . 
   
   
1 1 E[NC1,24 ]
The interpolated quadratic lines of best-fit for sectors ZMP22, ZMP28, and ZMP42 are

shown in Figure 19 and extrapolated up to the 3X traffic intensity. The extrapolated best-

fit lines only consider the 1X and subsampled traffic intensities they do not take into account

the conflict counts for the compressed traffic scenarios.

The extrapolated quadratic line of best-fit provides a basis to verify that the 24 hour

uncontrolled conflict totals can be estimated by scaling the compressed conflict totals (i.e.

E[NCI,24 ] = N̄CI,C × I for traffic intensities with I > 1). Figure 19 includes the estimated

24 hour uncontrolled conflict totals, E[NCI,24 ], according to the hypothesized scaling for the

sectors ZMP22, ZMP28, and ZMP42.

To assess the validity of the extrapolated expected results with the hypothesize of un-

controlled conflict totals, the coefficient of determination, R2 , of the model in Equation 5 is

calculated. In the general case, with observed values θ̂, and modeled values θ, the R2 value

is given by
− θi )2
P
2 i (θ̂i
R =1− P . (9)
i (θi − mean(θ̂))2
An R2 value close to 1 indicates a good fit. The R2 values for the average 24 hour scaled

uncontrolled conflict totals in each of the 22 sectors are provided in Figure 20 . In the

calculations, the observed values, θ̂, are given by scaling the uncontrolled conflict totals

N̄CI,24 to E[NCI,24 ] for the compressed traffic intensities according to Hypothesis 1. The

predicted number of uncontrolled conflict totals, θ, is provided by the extrapolated quadratic

line of best-fit from Equation 7 and Equation 8 using the sub-sampled traffic scenarios. For

all sectors, the R2 values approach 1. Therefore, the hypothesis that scaling the uncontrolled

35
 1

R2
0.5

0
11 12 13 15 16 17 18 19 20 22 23 24 25 28 29 30 38 39 40 42 43
ZMP Sector
Figure 20: Coefficient of determination, R2 for models based on sub-sampled traffic scenarios
used to predict conflict totals as a function of traffic intensity in each sector.

conflict totals according to Equation 5 is believed to be valid. Even for the worst-case

fit, corresponding to ZMP22, the R2 value is .9478. Later studies in this thesis extend

application of Hypothesis 1 to smaller time increments.

The quadratic model in Hypothesis 1 follows the relationship between traffic intensity

and uncontrolled conflicts also described in the study found in [48]. Furthermore, the

quadratic model is consistent with the analytical results in [85]. By maintaining similar

conflict models, aspects of the compressed traffic model for generating potential conflicts at

high-intensity traffic levels are verified.

2.7 Review

This chapter presents a modeling procedure to generate uncontrolled traffic scenarios based

on a seed day. The goal of the modeling procedure is to create uncontrolled traffic sce-

narios that contains the same variations present in current air traffic operations, yet that

match the original traffic pattern of the seed day without being overly specific. Hypoth-

esis testing demonstrated that it is likely both these goals were achieved (i.e. the null

hypothesis is never rejected) through the resampling of aircraft trajectories according to

origin-destination pairs. Ideally, by matching the spatial distribution of aircraft trajecto-

ries, the ultimate goal of creating traffic scenarios that manifest in potential conflicts similar

to real-world operations is achieved. Extending upon this idea, it is also desired that the

generated potential conflicts match the behaviors and properties at current traffic levels,

36
and that they also model potential conflicts at higher traffic intensities than have yet to be

observed in real-world operations. Key properties include: the number of aircraft involved,

the location of the potential conflicts, the flight phase of each aircraft, etc. If these and

other key properties are maintained, then it can be assumed that the overall experience of

the air traffic controller is preserved according to the air traffic model. While many of these

key properties cannot be verified, the overall modeling procedure suggests a satisfactory

model.

While other models ([3, 10]) exist for generating uncontrolled aircraft trajectories, the

model presented here is a first in creating traffic scenarios without flight plans, while at-

tempting to model possible variations and uncertainties through a data-driven model. It

should be noted, however, that there is limited scope in the relevance of the proposed pro-

cedure for generating traffic scenarios. Thus, the procedure may only be applicable to the

study at hand involving the characterization of the conflict-detection and resolution process.

37
 CHAPTER III

SECTOR CONFLICT ANALYSIS

To gain a better understanding of the traffic and the conflict event-process within a sector, an

initial study of the uncontrolled flight plans generated in Chapter 2 is provided. The analysis

in this chapter motivates the improvement of conflict-detection tools and the introduction

of conflict-resolution systems in air traffic control. The analysis focuses on the number

of potential conflicts and the rate at which potential conflicts occur for uncontrolled air

traffic over a single day. Additionally, a taskload utilization analysis is performed based on

the minimum required communication time between controllers and pilots. Ultimately, the

analysis framework and the results motivate the inclusion of unified conflict-detection and

resolution decision-support tools to aid air traffic controllers in managing and separating

aircraft within an airspace.

In current en route operations, a conflict occurs between two aircraft when they break

minimum separation requirements (5 NM laterally and 1000 ft vertically). When an air

traffic controller identifies a pair of aircraft that has the potential to break minimum sepa-

ration requirements, the event is labeled a potential conflict. For a potential conflict, the air

traffic controller should take action to prevent its occurrence. Thus, the number of poten-

tial conflicts arising in an airspace is one measure of controller’s effort . While there are a

number of other tasks that controllers must attend to in current day operations - including,

but not limited to, acknowledgements, hand-off, and clearances - the time spent resolving

potential conflicts can become taxing. One study estimates that for each potential conflict,

the controller spends 27.6 seconds identifying it, generating a resolution command, and

communicating the command to the appropriate aircraft [20]. More so, this time does not

include verification that the resolution command is properly implemented by the pilots and

aircraft. As traffic levels increase, air traffic controllers will likely dedicate greater amounts

of time and mental effort to handling potential conflicts. This chapter demonstrates that

38
if current practices are continued, then the effort required to separate aircraft will reach

a point when air traffic controllers are no longer able to maintain situation awareness and

ensure safety within the airspace.

To estimate the number of potential conflicts an air traffic controller resolves, the uncon-

trolled flight paths generated in the previous chapter are processed to identify uncontrolled

potential conflicts. An uncontrolled potential conflict is said to occur if any two aircraft
min NM and 1000 ft of each other; unless stated otherwise, D min = 9 NM
come within Dsep sep
min is not prescribed by the 5 NM en route regulation). The separation distance
(Note, Dsep
min , and used in the labeling of potential conflicts, is an approx-
criteria of 9 NM, set by Dsep

imate value related to controller action and concern [6, 7, 40, 101]. This is not to say a

dangerous situation will arise if aircraft come within 9 NM. Rather, the separation distance

reflects the fact that controllers behave conservatively in how they space aircraft. The extra

buffer distance beyond the 5 NM requirement is a result of cautious actions taken by air

traffic controllers to ensure clear separation between aircraft. Thus, if aircraft come within
min of each other, the model in this chapter assumes the air traffic controller issues a reso-
Dsep
min is henceforth referred to as the aircraft
lution command. Accordingly, the parameter Dsep

spacing distance.

According to the uncontrolled potential conflicts identified by the criteria above, an

undirected graph model is formed to represent the conflict relationships between all aircraft.

For a scenario representing a single day of traffic, the corresponding constructed conflict

graph is denoted by G = (V, E). Aircraft are represented by nodes in the vertex set V =

{n1 , . . . , nM }, where node ni corresponds to aircraft Ai . Potential conflicts are indicated by

an undirected edge in the edge set E. That is, for a potential conflict between aircraft Ai

and Aj , there is an edge (ni , nj ) ∈ E. An example conflict graph is provided in Figure 21.

For this chapter, the graph is assumed to be static and complete, implying that accurate
min = 9 NM aircraft spacing
knowledge of the uncontrolled flight plans exists up to the Dsep

distance. It is then assumed that all uncontrolled potential conflicts coming within 9 NM

are deemed to require a resolution command. By taking this approach, the graph represents

all aircraft passing through the airspace and all potential conflicts an air traffic controller

39
 n1 n5 n6
n7
n2
n3 n4 n8 n9

Figure 21: Example of conflict graph G with three conflict-clusters.

would attempt to resolve. A more formal model on the creation and definition of the conflict

graphs is presented in Chapter 4 and further extended to the dynamic case.

In relation to the provided description of conflict graphs, the number of uncontrolled

potential conflicts in a day, for an arbitrary airspace, is given by the total number of edges

in the edge set E.

Definition 1. For a conflict graph, G = (V, E), representing traffic for a predefined time

period and airspace according to the modeling description in Chapter 2, the number of un-

controlled potential conflicts is defined as card(E).

The analysis in this chapter also considers the complexity of potential conflicts. Here,

the complexity of a cluster of potential conflicts is associated with the number of potential

conflicts and the number of aircraft involved. Consider the examples of potential conflicts

provided in Figure 22. For each case, let aircraft be traveling on the same flight-level at

equal speeds. For the potential conflict between two aircraft illustrated in Figure 22a,

the conflict complexity is said to be low (e.g. similar aircraft speeds and no emergency

situations related to the aircraft). There is only one potential conflict and two aircraft.

Hence, the conflict-resolution problem is straight forward to solve barring any other difficult

circumstances outside the two aircraft (e.g. weather, airspace restrictions, or other aircraft).

In fact, the safety of a number of pairwise conflict-resolution algorithms has been formally

verified [26, 37, 61]. Figure 22b shows a set of 4 aircraft converging to a single point at

the same time. The conflict complexity for this configuration is considered to be higher; a

larger number of aircraft is involved, and all aircraft have potential to be in conflict with

one another. However, the number of aircraft does not necessarily imply a high level of

conflict complexity in solving the conflict-resolution problem. The example provided in

40
 A4 A1

A2
A3
A1 A1 A3 A4
A5
A2 A2 A6
(a) Simple poten- (b) A high complex- (c) A low complexity set of po-
tial conflict with ity set of potential con- tential conflicts with many air-
two aircraft. flicts with many air- craft.
craft.

Figure 22: Potential conflicts can be characterized by the number of aircraft involved and
overall complexity of generating a resolution.

Figure 22c depicts a single aircraft that has potential conflicts with 5 other aircraft. While

the number of aircraft in this example is greater than that in Figure 22b, with the same

number of potential conflicts, the conflict complexity is considered less. Resolution of all

potential conflicts is possible with a single aircraft maneuver. Provided sufficient time and

free airspace exist, a single altitude change to aircraft A1 is a possible resolution.

In general, the notion of conflict complexity is somewhat vague; while there is some

relationship between the number of aircraft involved as well as the number of potential con-

flicts, other factors such as geometry and variation in aircraft properties (e.g. groundspeed

or dynamic capabilities) are also important. Furthermore, conflict complexity is not only

intrinsic to the circumstances of the potential conflicts but also depends on the conflict-

resolution process. Poor decision making that creates secondary conflicts or overly complex

dynamic models can make the conflict-resolution problem difficult to solve, and possibly

intractable.

In this chapter, the assessment of conflict complexity is limited to a description of the

number of uncontrolled conflicts; the rate at which the uncontrolled conflicts occur; and

the average number of aircraft and potential conflicts involved in multi-aircraft conflicts.

Based on the air traffic model described in Chapter 2, analysis indicates that as the traffic

volume increases, the number of potential conflicts in an airspace and the average number

41
of potential conflicts per aircraft increase. Growing complexity supports the claim that air

traffic controllers require additional support through the improvement of conflict-detection

tools and the introduction of conflict-resolution decision-support tools.

Prior to the analysis, additional vocabulary is required to describe conflict complexity

of an air traffic scenario. Making use of the definition provided in [41], the term conflict

clusters is introduced.

Definition 2. A conflict cluster is a sub-graph formed by the transitive closure of aircraft

pairs in potential conflict. For example, if aircraft Ai and Aj have a potential conflict, and

Aj and Ak have a potential conflict, then Ai , Aj , and Ak , form part of a conflict cluster,

and define an associated sub-graph. [41]

For an example of conflict clusters, consider the arbitrary graph, G, illustrated in

Figure 21 representing a simple traffic scenario. The conflict graph contains 3 conflict-

clusters. Two conflict clusters, i.e. sub-graphs, represent aircraft that all have potential

conflicts with each other. In the third sub-graph, although there is no potential conflict

between aircraft A6 and A9 , represented by nodes n6 and n9 , they still form part of the

same conflict cluster.

3.1 Assumptions

This chapter analyzes the number of uncontrolled conflicts occurring in an airspace. It also

proposes a simple communication taskload model. The communication model is used to

determine the likelihood that events in an airspace require more attention than the air traffic

controller can provide within a given timeframe. Both these studies and their associated

models utilize the following assumptions:

• Potential conflicts affect the air traffic controller uniformly. Further, the amount of

effort required by the air traffic controller in spacing aircraft is uniform, regardless of

the conflict configuration. For example, the difficulty of resolving a potential conflict

between two cruising aircraft is assumed to be the same as managing a potential

conflict between an ascending aircraft and a cruising aircraft.

42
 • The number of potential conflicts is unaffected by the conflict-resolution process.

• Controller decisions to maneuver an aircraft are based on a deterministic model that

considers lateral separations between aircraft at the same flight-level. The value of
min is used to label potential conflicts as requiring resolution.
parameter Dsep

• Communication times between controllers and pilots for specific events are static,

regardless of situational factors within the airspace. Furthermore, events (e.g. hand-

off, potential conflicts) are handled by the controller in the order they occur.

• It is unsafe for air traffic controllers to sustain 10 minutes of continuous communication

and control of aircraft without respite.

The second assumption is removed in future chapters in the thesis. This allows for

comparison of advisory conflict-detection and resolution tools in reducing the effort related

to the conflict-resoluiton process.

3.2 Uncontrolled Conflicts within a Sector

In current operations, an air traffic controller’s priority is to maintain a safe airspace. Thus,

if the number of potential conflicts an air traffic controller is presented with regularly exceeds

his or her ability to resolve them in the required amount of time, then systematic change

is required. Otherwise, if air navigation service providers such as the Federal Aviation

Administration do not address the issue of overworked air traffic controllers, then unsafe

operating environments will become increasingly prevalent.

Following the traffic model presented in Chapter 2, Hypothesis 1, and the labeling of un-
sep
controlled conflicts utilizing the spacing parameter Dmin = 9 NM, analysis indicates that as

traffic intensity increases, the number of uncontrolled conflicts increases as well. Figure 23

illustrates the expected number of uncontrolled conflicts for the 100 traffic scenarios rep-

resenting June 14th for different traffic intensities. Considering all sectors in Minneapolis

Center (ZMP), ZMP42 contains the greatest average number of uncontrolled conflicts over

the 24 hour time period at each traffic intensity. The sector is located in the southern region

of the center and contains a portion of the airspace between FL350 and FL380. Because

43
 of Uncontrolled Conflicts
800 I=1X

Expected Number
I=2X
600 I=3X

400

200

0
11 12 13 15 16 17 18 19 20 22 23 24 25 28 29 30 38 39 40 42 43
Sector

Figure 23: Conflict totals for sectors in Minneapolis Center over a 24hr time period when
sep
Dmin = 9 NM.

(a) ZMP42 (b) ZMP12 (c) ZMP16

Figure 24: Traffic Density Maps.

of the sector’s spatial location and air traffic demand within the National Airspace, the

majority of traffic within ZMP42 is high altitude cruising aircraft traveling eastbound or

westbound. Figure 24a, a density plot of the traffic within the sector, illustrates the east-

bound and westbound nature of the traffic flow, as well as the 2D spatial traffic distribution

over the airspace. Two other airspaces of interest are sectors ZMP12 and ZMP16. These

two sectors contain air traffic patterns representative of distinct classes of airspace. In the

following analysis, the key differences between the sectors are highlighted.

The sector ZMP12 is similar to ZMP42, in that the majority of traffic is in its cruising

phase. However, there are numerous distinctions between ZMP12 and ZMP42. First, air

traffic within ZMP12 is more diverse in directionality, and according to the traffic density

map in Figure 24b, air traffic is relatively sparse outside of primary traffic flows that cover an

area 67% larger than ZMP42. Additionally, traffic in ZMP12 covers a larger range of flight

44
levels from FL240 and up and, as such, has a greater fraction of ascending and descending

traffic (∼ 37%) in comparison to ZMP42 (∼ 7%). The difference in the proportions of

ascending, descending, and cruising traffic is visible by flight phases distributions shown in

Figure 25a and Figure 26a. While 92.5% of all traffic in ZMP42 in its cruising phase, the

fraction of cruising aircraft in ZM12 is 63.1% of the traffic.

Unlike ZMP12 and ZMP42, sector ZMP16 is dominated by descending traffic (∼ 73%).

The sector is located near Minneapolis - St. Paul International Airport; as such, much of

the traffic in ZMP16 corresponds to arrivals coming from the east that are headed towards

the airport. Because of the dominant traffic pattern within the sector, as illustrated in

Figure 24c, the majority of uncontrolled conflicts occur at the merge point near the center

of the sector. Uncontrolled conflicts also occur along each of the arrival routes as aircraft

are lining up to enter the airport terminal area. Accordingly, the qualitative characteristics

of the uncontrolled conflicts in ZMP16 are different from those in ZMP12 and ZMP42. As

shown in Figure 25, Figure 27, and Figure 26, the distribution of ascending, descending

and cruising aircraft is different between the sectors, and likewise, so is the classification

of uncontrolled conflicts. In sector ZMP12, most aircraft are found in their cruising phase,

while a sizable remainder of flights are either ascending or descending at some point in

the sector. Accordingly, most of the uncontrolled conflicts involve at least one cruising

aircraft and smaller fractions involve two aircraft that contain a combination of ascending

and descending flights. In contrast, for ZMP42, which contains an even greater percentage

of cruising aircraft, approximately 60% of all uncontrolled conflicts involve two cruising

aircraft. Furthermore, as shown in Figure 26b, less than 1% of the uncontrolled conflicts

involve two aircraft that are ascending or descending, or a combination of the two.

The location of uncontrolled conflicts is also a reflection of the dominant traffic patterns

within the airspaces. As mentioned previously, and illustrated in Figure 28a, the majority

of uncontrolled conflicts in ZMP16 occurs at the merge point in the direction of a major

airport and along each of the descent routes. The traffic pattern in ZMP42 engenders un-

controlled conflicts along the dominant eastbound and westbound traffic routes. In contrast,

as shown in Figure 28b, with a larger fraction of crossing traffic, a significant portion of

45
 60 30

Percent

Percent
40 20
20 10
0 0
Cruise

Ascent

Descent

Cruise
Cruise
Cruise
Ascent
Cruise
Descent
Ascent
Ascent
Ascent
Descent
Descent
Descent
Flight Phase Conflict Configuration
(a) Aircraft flight phase distribution (b) Conflict configuration distribution

Figure 25: ZMP12 : Probability distribution of flight phases.

100
60
80
Percent

Percent

60 40
40
20
20
0 0
Cruise

Ascent

Descent

Cruise
Cruise

Ascent

Descent
Ascent
Ascent

Descent
Descent
Descent
Cruise

Ascent
Cruise

Flight Phase Conflict Configuration

(a) Aircraft flight phase distribution (b) Conflict configuration distribution

Figure 26: ZMP42 : Probability distribution of flight phases.

40
60
Percent

Percent

40
20
20
0 0
Cruise

Ascent

Descent

Cruise
Cruise

Ascent

Descent
Ascent
Ascent

Descent
Descent
Descent
Cruise

Ascent
Cruise

Flight Phase Conflict Configuration

(a) Aircraft flight phase distribution (b) Conflict configuration distribution

Figure 27: ZMP16 : Probability distribution of flight phases.

46
 #Descending aircraft≥1 #Cruising aircraft≥1 Two cruising aircraft
others others others
(a) ZMP16 (b) ZMP12 (c) ZMP42

Figure 28: Location of uncontrolled conflicts according to conflict configuration.

uncontrolled conflicts occurs at route crossings in ZMP12. This leaves large areas of the

airspace void of potential conflicts. The difference in routing and conflict manifestation

between ZMP12 and ZMP42 is also expressed through the distribution of aircraft crossing

angles for the uncontrolled conflicts. As shown in Figure 29a and Figure 29c, the uncon-

trolled conflict crossing angles for ZMP12 are more evenly dispersed than for ZMP42. A

larger fraction of uncontrolled conflicts occurs at high crossing angles (i.e., 45o − 135o ) for

ZMP12. Using communication entropy, H 0 = − pi log(pi ), as a measure of diversity for

P

the conflict angles, the three sectors in ascending order of entropy are ZMP16 (H 0 = 2.49),

ZMP42 (H 0 = 2.76), and ZMP12 (H 0 = 3.36); higher values of entropy indicate greater

diversity. Again, the distribution of the crossing angles in ZMP12 is a result of the mix-

ture of east-west and north-south traffic, while ZMP42 is dominated by east-west traffic.

The function that ZMP16 serves in providing a rough initial spacing of traffic landing at

Minneapolis-St. Paul Airport is apparent in the the crossing angles as well. The largest

bulk of conflict crossing angles is between 0o − 30o , with a large peak around 0o , which is a

consequence of the merging and spacing operations along dominant traffic flows within the

sector.

The variety in conflict configurations and flight phases demonstrates that for an advisory

conflict-resolution tool to be useful, it must be able to handle a diversity of problems. Simple

conflict-resolution algorithms that only consider level flight aircraft are inappropriate for

virtually all sectors of airspace. Likewise, conflict-detection systems must be capable of

47
 30 30 30

Probability

Probability

Probability
20 20 20
10 10 10
0 0 0
0 50 100 150 0 50 100 150 0 50 100 150
Crossing Angle: θ [degress] Crossing Angle: θ [degress] Crossing Angle: θ [degress]
(a) ZMP12 (b) ZMP16 (c) ZMP42

Figure 29: Discrete empirical probability distribution of uncontrolled conflict crossing angles
(5o increments).

predicting potential conflicts over a wide range of traffic scenarios.

The total number of conflicts is not completely representative of the conflict-event pro-

cess within each sector. In fact, the conflict-event process is non-homogeneous and varies

according to traffic demand over the day. As shown in Figure 30, during the early morn-

ing and late night, the change in the cumulative number of uncontrolled conflicts is small

in comparison to the time period between 10AM and 8PM. An initial assessment of the

relationship between aircraft arrivals and conflict events is possible by comparing their cu-

mulative counts. As shown in Figure 31, the hours between 10AM and 8PM account for the

greatest traffic throughput in the sectors, which thereby drives the conflict-event process

during the same time period. When analyzing the effort required by air traffic controllers

to separate aircraft, these peak hours are most relevant; outside these times, air traffic

controllers are unlikely to exceed their workload limits. This result is further supported by

the communication analysis in Section 3.3.

Considering the 10 hour time window between 10AM and 8PM for each sector, it is

asserted that the number of uncontrolled conflicts is heavily dictated by the traffic intensity

and aircraft spacing distances. Applying and extending the assertion from Hypothesis 1 in

Chapter 2 to the 10 hour time window, the following hypothesis is made:

Hypothesis 2. The expected number of uncontrolled conflicts in a sector with traffic in-
min , and time window w, denoted by E[N I,w ] can be estimated
tensity I, spacing distance Dsep C

48
 of Uncontrolled Conflicts
ZMP12: Average
Cumulative Number

100 ZMP12: 15%-85% Percentile

ZMP16: Average
ZMP16: 15%-85% Percentile
ZMP42: Average
50 ZMP42: 15%-85% Percentile

0
0 2 4 6 8 10 12 14 16 18 20 22 24
Time of Day [HR]

Figure 30: Cumulative number of uncontrolled conflicts for sectors ZMP12, ZMP17, and
sep
ZMP42, when Dmin = 9 NM.

ZMP12: Average
800
Cumulative Number
of Aircraft Arrivals

ZMP12: 15%-85% Percentile

ZMP16: Average
600
ZMP16: 15%-85% Percentile
ZMP42: Average
400
ZMP42: 15%-85% Percentile

200

0
0 2 4 6 8 10 12 14 16 18 20 22 24
Time of Day [HR]

Figure 31: Cumulative number of aircraft arrivals for sectors ZMP12, ZMP17, and ZMP42.

49
 Table 1: Best-fit coefficients for the uncontrolled conflict models.

Sector c1 c2 c3 c4 R2 (Complete) R2 (Simple)

ZMP12 0.0755 1.6688 0.4186 0.0062 .9997 .9925
ZMP16 0.1258 1.3224 0.2564 -0.0313 .9998 .9857
ZMP42 0.1722 5.8083 -0.8882 0.0403 .9999 .9953

by a function of the form

E[NCI,w ] = c1 I 2 (Dsep
min 2
) + c2 I 2 Dsep
min min
+ c3 IDsep min 2
+ c4 I(Dsep ) . (10)

The form of Equation 10 is a conjecture supported through testing. Further, through

qualitatively analysis of real-world operations, there should be an understanding that if

min = 0 NM)
there is no traffic (i.e. I = 0) or there are no separation requirements (i.e. Dsep
min should be
then, inherently, there cannot be any conflicts. Thus, the variables I and Dsep

found in each term so that if either is value 0, then E[NCI,w ] = 0.

For the 10 hour time window considered, the coefficients (c1 , c2 , c3 , and c4 ), calculated

according to the least-squares best-fit of Equation 10, for sectors ZMP12, ZMP16, and

ZMP42 are provided in Table 1. Also included in Table 1 are the R2 for each line of best-fit

when the complete model with all coefficients is used.

When considering the relative influence each coefficient plays, the term I 2 Dsep
min is most

important in determining the expected number of uncontrolled conflicts; a best-fit with only

the term I 2 Dsep

min results in an R2 value of about .99 for each sector. The R2 values for the

simple model only considering the term I 2 Dsep

min (s.t. c = 0, c = 0,c = 0) is included
1 3 4

in Table 1. The remaining terms, I 2 (Dsep

min )2 , ID min and I(D min )2 , improve the model
sep sep

marginally. The major implication of such a result is that the number of uncontrolled

conflicts grows quadratically as traffic intensity increases. Furthermore, the number of

uncontrolled conflicts is linearly proportional to the aircraft spacing distance. Example

min for the
lines of best-fit for ZMP42 are shown in Figure 32 across different values of Dsep

simple model.

The relationship between traffic intensity and uncontrolled conflicts is supported by a

similar study in [48]. Barring increased regulation of transport by government agencies,

the traffic intensity is effectively an uncontrollable variable. Conversely, aircraft spacing

50
 min
Dsep = 3 NM

of Conflicts [10-20HR]
600 min
Dsep = 4 NM

Average Number
min
Dsep = 5 NM
400 min
Dsep = 6 NM
min
Dsep = 7 NM
min
Dsep = 8 NM
200
min
Dsep = 9 NM
min
Dsep = 10 NM
0
0.5 1 1.5 2 2.5 3
Traffic Intensity

Figure 32: Simulated number of uncontrolled conflicts and best-fit lines according to traffic
intensity and separation distance for ZMP42.

distances can be reduced through improvements in technology that promote changes in

policy.

Key technological improvements to support the reduction of spacing distances include

increased accuracy in trajectory prediction and aircraft position sensing. Reducing aircraft

spacing provides the opportunity to reduce the number of labeled potential conflicts re-

quiring resolution. Otherwise, if air traffic controllers continue to space aircraft at 9 NM,

and perform their tasks according to current methods, then a tripling in traffic intensity

will results in approximately an 850% increase in the expected uncontrolled conflict rate for

ZMP42 (700% and 800% increase for ZMP12 and ZMP16, respectively).

To maintain a similar number of uncontrolled conflicts at the 3X traffic intensity when

compared to current uncontrolled conflict counts, air traffic controllers in ZMP42 would

be required to space aircraft at 1.2 NM (1.5 NM for both sectors ZMP12 and ZMP16).

The possibility of spacing aircraft at these distances is hindered by a number of factors.

First, conflict-resolution tools, human air traffic controllers included, are required to issue

resolution commands minutes in advance of a potential conflict to ensure smooth and phys-

ically realizable trajectories for maneuvering aircraft. Otherwise, if resolution commands

are issued seconds before a potential conflict, the range of solutions that are dynamically

acceptable shrink significantly. Because of the required buffer time between issuing a reso-

lution command and when the potential conflict may occur, the uncertainty in predicting

aircraft trajectories must decrease significantly to allow for 1.2 NM separation. For current

51
 10 I = 1X 10 10

[Conflicts/5min]

[Conflicts/5min]

[Conflicts/5min]
I = 1.5X
I = 2X
5 I = 2.5X 5 5
I = 3X

0 0 0
0 10 20 0 10 20 0 10 20
Time of Day [HR] Time of Day [HR] Time of Day [HR]
(a) ZMP12 (b) ZMP16 (c) ZMP42

min = 9 NM.
Figure 33: Average rate of uncontrolled conflicts when Dsep

systems, estimates indicate that uncertainty in the along track position of aircraft grows at

a rate between .20-.25 NM/minute [47, 74, 103] at the 1 standard deviation level. Thus,

over 5 minutes, the 1.2 NM separation distance is almost eclipsed by the uncertainty in pre-

dicting the aircraft positions. The difficulty of predicting aircraft trajectories is exacerbated

when aircraft are turning or changing altitude.

However, even if it were technically possible to predict aircraft trajectories minutes in

advance, spacing aircraft at 1.2 NM is potentially unsafe due to the wake vortexes shedding

off aircraft. These wake vortexes can interfere with the stability of other nearby aircraft [99].

As it stands, aircraft at airports are required to maintain spacing separations for arrival

and departure procedures to prevent unsafe operating conditions due to wake turbulence

[44]. Also preventing smaller aircraft spacings is the ever-present safety margin needed in

case of aircraft or system failures. For example if aircraft position sensing systems go down

(i.e. ADS-B, or radar) or aircraft lose controllability through the loss of a rudder, ailerons,

or engine, then small separation distances hinder the ability of controllers and nearby pilots

to respond in time to prevent a potentially catastrophic situation in the case of converging

trajectories.

Assuming that the number of uncontrolled conflicts constitutes a significant portion of

the effort required by air traffic controllers in managing an airspace, then the rate that

uncontrolled conflicts occur is of potential concern as traffic intensity increases. Shown in

Figure 33, for sectors ZMP12 and ZMP16, the rate of uncontrolled conflicts remains below

52
 0.4 0.4 0.4

Probability

Probability

Probability
0.3 0.3 0.3
0.2 0.2 0.2
0.1 0.1 0.1
0 0 0
0 10 20 0 10 20 0 10 20
Number of Conflicts Number of Conflicts Number of Conflicts
(a) 1X Traffic (b) 2X Traffic (c) 3X Traffic

Figure 34: Cumulative distribution of the number of uncontrolled conflicts for the busiest
min = 9 NM for ZMP12.
5 minute period of the day when Dsep

0.4 0.4 0.4

Probability

Probability

Probability
0.2 0.2 0.2

0 0 0
0 10 20 0 10 20 0 10 20
Number of Conflicts Number of Conflicts Number of Conflicts
(a) 1X Traffic (b) 2X Traffic (c) 3X Traffic

Figure 35: Cumulative distribution of the number of uncontrolled conflicts for the busiest
min = 9 NM for ZMP16.
5 minute period of the day when Dsep

5 per 5 minute time period. For ZMP42, the average uncontrolled conflict rate reaches

an average maximum of 10 conflicts per 5 minutes at the 3X traffic intensity. Assuming

that air traffic controllers require 30 seconds to resolve each potential conflict, and little

follow-up or conformance checking is required, then the average conflict rates are still within

required limits to provide sufficient time for air traffic controllers to space aircraft. Further,

when one considers that air traffic controllers often employ quick and simple resolution

commands during heavy traffic loads [52], thereby reducing the time required to resolve

each potential conflict, then even at 3X traffic intensity, the average conflict rates may

appear to be manageable by air traffic controllers, when excluding other tasks (aircraft

acknowledgements, hand-off, clearances, etc.).

Perhaps more telling to future concerns is the distribution of the number of uncontrolled

conflicts within the busiest 5 minute time period over each of the 100 traffic scenarios. This

distribution provides information about the conflict-resolution taskload during the busiest

53
 0.4 0.4 0.4

Probability

Probability

Probability
0.3 0.3 0.3
0.2 0.2 0.2
0.1 0.1 0.1
0 0 0
0 10 20 0 10 20 0 10 20
Number of Conflicts Number of Conflicts Number of Conflicts
(a) 1X Traffic (b) 2X Traffic (c) 3X Traffic

Figure 36: Distribution of the number of uncontrolled conflicts for the busiest 5 minute
min = 9 NM for ZMP42.
period of the day when Dsep

time period of a given day. For sectors ZMP12 and ZMP16, as the traffic intensities increases

the mean and variance of the number of uncontrolled conflicts in the busiest 5 minutes

increase as well. These results are illustrated in Figure 34 and Figure 35 and are indicated

by the rightward shifting and increased width of the probability distribution functions.

For both sectors, at the 3X traffic intensity, there is a measurable probability that 10

or more uncontrolled conflicts will occur during the 5 minute period. Because of the air

traffic patterns within the sectors that bring about merging and spacing operations, the

large fraction of ascending and descending traffic, and the high angles of intersection, the

complexity and controller workload of resolving such conflicts becomes increasingly difficult

for the unaided air traffic controller [58]. Assuming an ability to resolve each conflict in

30 seconds, then the case of encountering 10 conflicts in 5 minutes leaves no time for the

air traffic controller to complete any other tasks - even those that maintain situational

awareness. Taken together, not only is the number of potential conflicts large, but so is the

difficulty of resolving each one.

Figure 36 shows distribution of the number of uncontrolled conflicts in ZMP42 for the

1X, 2X, and 3X traffic intensities. The number of uncontrolled conflicts in ZMP42 at 3X

traffic intensity is quite troubling. For approximately 20% of the traffic scenarios, during the

busiest 5 minute time period, 20 or more uncontrolled conflicts occur. If a scenario like this

manifested in real-world operations, then the air traffic controller would be required to issue

a resolution command, on average, every 15 seconds for a sustained period of time. Such

54
 1 1 1

P (Nac < nac )

P (Nac < nac )

P (Nac < nac )

0.8 I = 1X 0.8 0.8
0.6 I = 1.5X 0.6 0.6
0.4 I = 2X 0.4 0.4
I = 2.5X
0.2 0.2 0.2
I = 3X
0 0 0
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
Number of aircraft: nac Number of aircraft: nac Number of aircraft: nac
(a) ZMP12 (b) ZMP16 (c) ZMP42

Figure 37: Distribution of the number of aircraft invovled in each conflict cluster as traffic
min = 9 NM for ZMP42.
intensity increases for Dsep

1 1 1
P (Ne < ne )

P (Ne < ne )

P (Ne < ne )
0.8 I = 1X 0.8 0.8
0.6 I = 1.5X 0.6 0.6
0.4 I = 2X 0.4 0.4
I = 2.5X
0.2 0.2 0.2
I = 3X
0 0 0
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
Number of conflicts: ne Number of conflicts: ne Number of conflicts: ne
(a) ZMP12 (b) ZMP16 (c) ZMP42

Figure 38: Distribution of the number of edges in each conflict cluster as traffic intensity
min = 9 NM for ZMP42.
increases for Dsep

a sustained taskload would probably be considered inconceivable by air traffic controllers.

Even at the 2X traffic intensity, the number of uncontrolled conflicts in ZMP42 is quite

large. In fact, the distribution of uncontrolled conflicts is similar to ZMP12 and ZMP16 at

the 3X traffic intensity. This result indicates that if a uniform growth in air traffic demand

occurs, then ZMP42 will be impacted much earlier than other airspaces.

As traffic intensity increases, not only will the number and rate of potential conflicts

grow, but so will the complexity. Analysis of the traffic scenarios for each sector indicates

that with growing traffic intensity, the number of aircraft and number of uncontrolled con-

flicts associated with each conflict cluster increases. The sampled cumulative distributions

for both measures are provided in Figure 37 and Figure 38. The sampled distributions

for the sectors show that both the mean and variance for both measures of complexity

55
 1

Average Fraction of
Pairwise Conflicts
0.8
0.6
ZMP12
0.4
ZMP16
0.2 ZMP42
Others
0
0 0.5 1 1.5 2 2.5 3
Traffic Intensity

Figure 39: Fraction of pairwise uncontrolled conflicts.

(number of aircraft and number of conflicts) increase with traffic intensity (indicated by

the movement of the cumulative probability distributions towards larger values). From the

sampled distributions, more concise information can be extracted, such as the fraction of

conflict clusters that only include two aircraft (i.e. strictly pairwise conflicts). As shown

in Figure 39, as traffic intensity grows, the fraction of pairwise conflicts decreases. Such

a result implies that with multiplying traffic intensity, the ability of humans and conflict-

resolution tools to remain applicable while still using pairwise decision-making strategies

becomes increasingly limited. Additionally, with larger conflict clusters, the secondary ef-

fects of a single resolution maneuver become increasingly important. Conflict-resolution

should consider more than just the two aircraft that are part of the immediately pending

potential conflict. Other aircraft that form part of the same conflict cluster can be affected.

Despite the growth in conflict complexity with regards to the number of aircraft and

number of conflicts involved, there are indications that conflict complexity growth is only

weakly “quadratic,” especially when compared to the growth rate of uncontrolled conflicts.

That is, while the total number of uncontrolled conflicts in a sector grows quadratically, the

conflict complexity for each cluster grows at a much slower rate. Consider Figure 40a and

Figure 40b, which depict the average number of aircraft and average number of uncontrolled

conflicts per cluster. As traffic intensity increases, the growth in the number of aircraft and

uncontrolled conflicts is approximately linear. More interesting, is that when plotting the

number of aircraft per cluster versus the number of conflicts, their relationship is linear, as

56
 3.5 2.5
ZMP12

Conflicts per Cluster

Aircraft per Cluster
Average Number of

Average Number of
ZMP16
3 ZMP42 2
Others

2.5 1.5

2 1
0 1 2 3 0 1 2 3
Traffic Intensity Traffic Intensity
(a) Average number of aircraft per conflict (b) Average number of conflicts per conflict
cluster, with quadric line of best fit. cluster, with quadric line of best fit.

Figure 40: Average growth of conflict clusters as intensity increases.

Table 2: Coefficients and R2 values for least-squares lines of best-fit considering the ratio
of conflict complexity.

Sector Slope R2
ZMP12 1.135 .9997
ZMP16 1.300 .9977
ZMP42 1.218 .9995

shown in Figure 41. Table 2 contains the slope of the least-squares linear fit, along with

the coefficient of determination value, R2 , associated with each linear fit. The coefficient of

determination values indicates a strong fit with the linear model for the three sectors.

The relative rate of growth between the two complexity measures, i.e. the slopes pro-

vided in Table 2, is of particular interest because it provides a normalized measure related

to the amount of effort required to deconflict clusters as the cluster sizes grow. Consider

the two conflict clusters illustrated in Figure 42. In order to completely resolve all potential

conflicts, the first conflict cluster requires at least 2 maneuvers. Meanwhile, the second con-

flict cluster can be resolved with a single resolution command issued to the aircraft A5 . The

first conflict cluster corresponds to a special type of graph referred to as a complete graph;

complete graphs have edges connecting each node. For a complete graph with n nodes, the

total number of edges is n(n − 1)/2 [72], and accordingly, requires n(n − 1)/2 − 1 resolution

commands to prevent all potential conflicts. For the conflict-resolution problem requiring

57
 2.5

Conflicts per Cluster

Average Number of
2

ZMP12
1.5
ZMP16
ZMP42
Others
1
2 2.5 3
Average Number of
Aircraft per Cluster

Figure 41: Number of expected conflicts versus number of expected aircraft in each cluster.

n1 n6
n5
n2
n3 n4

Figure 42: Two conflict clusters of differing conflict complexity.

that all potential conflicts are resolved (i.e. edges removed from the graphs), conflict clus-

ters associated with complete graphs require equal or more resolution commands than any

other type of conflict cluster with the same number of nodes. Therefore, because the lines

of best fit in Figure 41 grow linearly, and not according to n2 , there is indication that the

conflict clusters are not exclusively complete graphs (outside of pairwise conflicts). In the

case where a non-complete conflict graph is present, a conflict-resolution program might be

able to appropriately select which aircraft to maneuver in order to reduce to the number of

commands used to space aircraft. For example, maneuvering aircraft A5 , instead of both

A4 and A6 , resolves all potential conflicts. The selection of which aircraft to maneuver to

reduce the conflict-resolution taskload is considered beginning in Chapter 5.

The linear fit between between the number of aircraft and the number of conflicts

dictates that on average, every additional aircraft in a conflict cluster is associated with

just over one potential conflict; a larger slope signifies faster complexity growth. For the

three sectors considered, the conflict complexity ratio of ZMP16 is the greatest, followed by

58
ZMP42, and ZMP12.

3.3 Controller Communication and Control Time

Despite the prior section’s emphasis on the conflict counts as a measure of difficulty, a more

appropriate estimate of controller overload is required. A better metric is needed to deter-

mine when controllers encounter scenarios that are beyond their capabilities. The process

described here in detecting unmanageable traffic levels is based on a taskload and commu-

nications model that is associated with the control process between air traffic controllers

and aircraft.

As it stands, the capacity to properly manage and separate air traffic directly depends

on the controller workload [11]. Unfortunately, controller workload is difficult to measure

quantitatively and depends on each individual controller’s capability and perception. In

current operations, instead of relying on controller workload as a measure for sector ca-

pacities, the Federal Aviation Administration has relied on a simple proxy: the number of

aircraft present in a sector. The limit on this value is established by the Monitor Alert Pa-

rameter (MAP). If aircraft counts are within the MAP value, then it is assumed that traffic

conditions are within the controller’s abilities. However, MAP values do not accurately rep-

resent sector capacity - and often times lead to congestion, or conversely, under-utilization

of the airspace because they does not accurately address performance limits.

Radio communication time has been considered an objective metric to evaluate con-

troller workload while managing traffic. A series of experiments have concluded that re-

alistic radio activities can be used to provide objective measures of workload [14, 21, 87].

Additionally, other studies have demonstrated the high correlation between communication

duration and controller workload, thereby effectively validating communication time as an-

other workload measure [63, 77]. By its very nature, communication time can be related to

bandwidth limitations within the human-in-the-loop control system. Because communica-

tion is used in both the management and control of aircraft, if events within the airspace

require greater amounts of communication than time permits, then some fundamental limit

59
has been reached. Accordingly, this section, and the proposed communication model, stud-

ies how often communication requirements exceed reasonable capabilities of the air traffic

controller.

In current operations, each aircraft passing through a sector is communicated with at

least twice by the managing air traffic controller: once to acknowledge the aircraft as it

enters the sector and again when the aircraft is handed-off to the next sector. Another

prevalent communication type typically occurs when an airspace is congested and there is

the potential for conflict. Provided sufficient concern by air traffic controllers exists that a

pair of aircraft might conflict, then a resolution command is issued. In the case of a potential

conflict, air traffic controllers must determine safe routes for all aircraft and communicate

them to each pilot. For this process to occur in a safe manner, there must also be sufficient

time for the controller to gain situational awareness and to monitor conformance of the

resolution commands. Finally (as part of the current system of clearance based control, in

which requests are made by pilots, and verified for safety by the controller), any request for

changes in heading, speed or altitude requires communication. The most common of these

pilot requests is for altitude changes, either to ascend or descend in flight-level.

Using the 100 traffic scenarios created in Chapter 2, an event process for each scenario

is generated. The set of events considered as part of each scenario’s event process includes

entering and exiting aircraft, as well as the detection and resolution of potential conflicts.

Further, entering aircraft are classified according to whether or not they require clearances

for altitude changes. Each of the events gives rise to a response and communication se-

quence between the controller and the pilots. By summing up the communication time, a

measure of the required control effort is generated. However, in this case, because some

communication events are not accounted for (e.g. read-backs, and weather and congestion

advisories), the estimated communication times represent a lower bound, when assuming

static communication times.

The event process for a traffic scenario is represented by the function fevent (t), which

contains a sequence of Dirac delta functions. An example depiction of fevent (t) is illustrated

in Figure 43. The magnitude of each Dirac delta function corresponds to the amount of

60
 Ascending arrival Event Time Line
Cruising arrival Tw
Descending arrival
⋆
Exiting aircraft
fevent (t) t t=0
Potential conflict
Wwindow (t)
Figure 43: Representation of the event process within an airspace, and method for calcu-
lating the minimum estimated controller communication time.

time and effort required to communicate commands and information between the controller

and the pilots for that event. For example, when an aircraft enters an airspace and requests

to ascend in altitude, the communication time from the initial acknowledgement to the

granted clearance takes approximately 11 seconds - that is, if the air traffic controller does

not detect any immediately pending conflicts as a result of the maneuver. Table 3 contains

the communication times associated with each event considered in this communications

model. Except for potential conflicts, the communication times are averages based on a

sample audio recording of air traffic controllers within the Atlanta Center on November 3,

2010 over a 1 hour time period. The communication time for potential conflicts is selected

to be 20 seconds. While the time is less than the previously cited 27.6 seconds [20], it is

representative of more efficient actions taken by the air traffic controller. In this way, the

resulting model for channel utilization more closely approximates a lower bound on real-

world practices. Denoting the sequence of times for arriving level aircraft by {tarr,l
n }; arriving

descending aircraft by {tarr,d

n }; arriving ascending aircraft by {tarr,a
n }; exiting aircraft by

{texit con
n }; and potential conflicts by {tn }; the complete event process function for a traffic

scenario is

(δ(t − tarr,l (δ(t − tarr,d

X X
fevent (t) =W arr,l i )) + W arr,d i ))
i i

tarr,a
X X
+ W arr,a (δ(t − i )) + W exit (δ(t − texit
i )) (11)
i i
X
+ W con (δ(t − tcon
i )),
i

where the weightings W arr,l , W arr,d , W arr,a , W exit , and W con correspond to the times in

61
 Table 3: Average communication time for each event.

Event Average Time [seconds] Weighting Variable

Arrival at level-flight 6 W arr,l
Arrival requiring clearance 11 W arr,d and W arr,a
Departure 6 W exit
Potential conflict 15 W con

Table 3.

To calculate the communication time over all time intervals for an entire traffic scenario,

the event process function fevent (t) is convolved with a square-function centered at 0 with

the appropriate time-width. That is, for a time interval of Tw minutes, the square-function

is 
1, if t < Tw


hsq (t) = (12)


0 else.

The resulting communication time, CTw (t), within Tw /2 minutes of time t is

CTw (t) = hsq (t) ? fevent (t). (13)

An illustration depicting the computation of communication times within a desired time

interval is shown in Figure 43.

The percent utilization of the communication channel for the time window Tw , denoted

UTw (t), is calculated by the equation

UTw (t) = CTw (t)/Tw . (14)

If at time t, UTw (t) = 1, then events within the airspace require that the air traffic controller

and pilots communicate for the complete Tw minutes. Note, the case when UTw (t) > 1

does not necessarily imply infeasibility in regards to the communication and control limits.

In practice, some communication events can be processed earlier (e.g hand-offs) or later

(e.g. acknowledgements, and clearances) than normally would occur without loss of safety

or situational awareness. Thus, time-shifting of events can still allow for feasibility over

shorter time windows. However, while UTw =2.5 (t) > 1 and UTw =5.0 (t) > 1 might still be

feasible, UTw =10.0 (t) > 1 is surely not.

62
 I = 1.0X I = 1.0X
1.5 I = 2.0X 1.5 I = 2.0X
I = 3.0X I = 3.0X
1 1

0.5 0.5

0 0
0 5 10 15 20 0 5 10 15 20
(a) Tw = 2.50 minutes (b) Tw = 5.00 minutes

I = 1.0X I = 1.0X
1.5 I = 2.0X 1.5 I = 2.0X
I = 3.0X I = 3.0X
1 1

0.5 0.5

0 0
0 5 10 15 20 0 5 10 15 20
(c) Tw = 10.00 minutes (d) Tw = 20.00 minutes

Figure 44: Average utilization time with upper and lower quartiles for ZMP42.

The expected utilization times, E[UTw (t)], for ZMP42 for different values of Tw and

traffic intensities are shown in Figure 44. Similar to conflict counts within the sector, the

expected utilization time increases significantly during peak traffic hours. Furthermore, as

traffic intensity increases, the average utilization time multiples. The utilization times for

ZMP12 and ZMP16 with Tw = 5 minutes are illustrated in Figure 45.

While it is tempting to use the utilization times UTw (t) to estimate controller workload

or to establish when traffic conditions exceed reasonable workload limits, such an approach

I = 1.0X I = 1.0X
1.5 I = 2.0X 1.5 I = 2.0X
I = 3.0X I = 3.0X
1 1

0.5 0.5

0 0
0 5 10 15 20 0 5 10 15 20
(a) ZMP12 (b) ZMP16

Figure 45: Average utilization time with upper and lower quartiles for Tw = 5 minutes.

63
is not taken. Because the manifestation and dictation of workload limits in each of the

sectors and for each controller are different, it is not possible to establish a consistent

bound. As noted in prior results in this chapter, the type of potential conflicts commonly

encountered varies significantly between the sectors. Because the effort required to resolve

different types of conflicts is not uniform, a simplistic model that linearly scales and sums

the communication times does not accurately capture controller workload. Furthermore,

in human-factors research there is a recognition that “workload is not something imposed

upon a passive [air traffic controller] but, rather, is something the [air traffic controller]

actively manages” [59]. In virtually all aspects of their role, air traffic controllers adjust

their strategies in response to system changes, future occurrences, and perceived workload.

Instead of using UTw (t) as a measure of controller workload, the utilization time is

checked to see if it ever exceeds the value of 1. In this manner a more robust and conser-

vative approach is taken to understanding the limits of air traffic controllers. And again,

there should be an understanding that if events within an airspace require more time than

is available, then no matter what strategies the controller takes in regulating his or her

workload, proper management of the airspace is not possible.

Following the analysis framework above, the probability that utilization rates exceed the

capability of the air traffic controller at some point during the day, i.e. P (∃t|UTw (t) > Tw ),

is considered. Excessive utilization is a suitable measure to gauge when traffic conditions

are no longer manageable – likely well beyond human controller workload limits. Figure 46

shows the cumulative probability distributions of excessive utilization for different values

of Tw . According to the results in Figure 46c, for the 5 minute time window, at 1.7X

traffic levels in ZMP42, the minimum required communication and control effort exceeds

the time allowed approximately 50% of the time. For the sectors ZMP12 and ZMP16, the

corresponding traffic levels when utilization rates at Tw = 5 minutes are broken 50% of

the time occurs at the 2.6X and 3X traffic intensities, respectively. The results suggest

that these traffic intensities, while potentially manageable, are risky. This conclusion is

supported by the human-in-the-loop simulations in [54] that indicated controllers struggle

to maintain safe operating conditions at the 2X traffic level, and fail at the 3X traffic level.

64
 1 1

P (UTw (t) > Tw )

P (UTw (t) > Tw )

0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
1 1.5 2 2.5 3 1 1.5 2 2.5 3
Traffic Intensity Traffic Intensity
(a) ZMP12 (b) ZMP16

1
P (UTw (t) > Tw )

0.8 Tw = 2.5 min

0.6 Tw = 5.0 min
0.4 Tw = 10.0 min
0.2 Tw = 20.0 min
0
1 1.5 2 2.5 3
Traffic Intensity
(c) ZMP42

Figure 46: Probability of over-utilization for different time-windows, Tw .

Working under the assumption that air traffic controllers are unable to manage sustained

efforts, perhaps equally important is identifying traffic levels when P (∃t|UTw =10 (t) > 1).

That is, air traffic controllers are presented with scenarios that require more than 10 con-

tinuous minutes of communication and control time. For the sectors ZMP12, ZMP16, and

ZMP42, breaking medium-term communication and control limits occurs at the 2.25X,

2.25X, and 1.5X traffic intensities. While more restrictive than the 50% limit set above,

these margins are more reflective of the longer-term management of airspaces. Even at these

traffic levels, there exists the probability that within 5 minute time windows, communication

and control limits are exceeded.

3.4 Review and Analysis

This chapter explores the uncontrolled conflict event process in sectors as traffic intensity

increases. Assuming that air traffic control continues with the same policies, practices,

and technologies in current operations, the forecast models suggest that increased traffic

intensities will ultimately lead to traffic scenarios that are beyond the capabilities of the

65
unaided air traffic controller. Of particular concern is the quadratic growth rate of uncon-

trolled conflicts in relation to traffic intensity. For a doubling in traffic, the conflict count

model proposed in this chapter predicts that uncontrolled conflicts will quadruple. And for

sectors that are already heavily sectorized, more radical changes beyond airspace redesigns

are required if a doubling of potential conflicts is to be managed by the human controller.

From an engineering stand-point, traffic demand is an uncontrollable variable that will

continue to grow in the coming decades. To ameliorate problems associated with increased

traffic levels, it is conjectured that there is a pending need to aid the air traffic con-

troller through improved conflict-detection and conflict-resolution decision-support tools.

Improved conflict-detection systems require improved trajectory prediction and position

sensing of aircraft. Thus, while traffic intensity cannot be controlled, efforts to improve

trajectory prediction and position sensing systems will likely enable a reduction in air-
min . Results in this chapter indicate that by reducing aircraft
craft spacing distances, Dsep

spacing, air traffic controllers can make use of improved trajectory prediction tools to re-

duce the number of potential conflicts that need resolution commands. With a reduction

in the excessive labeling of potential conflicts, the conflict-resolution taskload will surely

decrease. However, there are limits to which decreased separation standards can aid the

air traffic controller. Because of the potential for failure of position-reporting systems and

the ever-present error in trajectory prediction (even if uncertainty is reduced), aircraft are

likely to continue to be spaced at greater distances then necessary. Continuing onward,

conflict-resolution tools may be able to supplement the gains made by reducing aircraft

spacing.

One primary advantage of advisory conflict-resolution tools is that they transfer part of

the mental taskload from the human controller to a computer system. Such transference is

not only important in mitigating the workload implications of increasing conflict-resolution

taskload, but also because potential conflicts will become increasingly complex with greater

traffic intensity. Potential conflict-clusters will evolve to contain greater numbers of air-

craft and interactions. While previously the conflict-resolution process could be solved in

66
a sequential pairwise fashion with little worry, greater traffic loads generating more com-

plex conflict inactions will inhibit this strategy. Computer decision-support tools have the

potential to play a key role in resolving these more complex conflict scenarios. However,

the design of conflict-resolution algorithms to form the foundation of an advisory system

should not be considered straight forward. As demonstrated in this chapter, any conflict-

resolution tool will be presented with handling a wide arrays of problems, conflict types (e.g

ascending/descending aircraft, mixed aircraft models and capabilities) and geometries.

In summary, this chapter has indicated a great need for improvement in the polices,

practices, and technologies used in current day air traffic control. Barring any adjustments,

the human controller will be overwhelmed with the requirements of managing and separating

aircraft in an airspace. Thus, there is a strong imperative to improve conflict-detection

systems and introduce advisory conflict-resolution tools into the repertoire of air traffic

controllers.

67
 CHAPTER IV

GRAPH REPRESENTATIONS OF POTENTIAL CONFLICTS

Through simulations of uncontrolled air traffic at high intensities, the previous chapter rev-

eled that the rate of potential conflicts, in addition to other air traffic control events (e.g.

acknowledgements, clearances), will one day hinder the ability of the air traffic controller

in maintaining situational awareness and safety. To prevent unsafe operating conditions,

air navigation service providers are taking action by developing next generation conceptual

systems such as NextGen and SESAR. These next generation systems adjust policies and

practices, and integrate new technologies. Researchers and air navigation service providers

hypothesize that these adjustments will aid air traffic controllers by reducing and trans-

forming controller workload, thereby enabling increases in airspace capacity to support

future air traffic demand. While there are numerous components to these next genera-

tion conceptual frameworks, semi- and fully-automated conflict-detection and resolution

systems is a promising component. This chapter begins to explore the design, capability,

and implementation of conflict-detection and resolution systems to better understand their

interactions with human air traffic controllers in an advisory human-in-the-loop framework.

As stated previously, the metric of interest for this thesis is the taskload associated with

the conflict-resolution process, specifically, the number of advisory resolution commands an

air traffic controller issues when acting in conjunction with an advisory conflict-detection

and resolution system.

Addressing controller taskload is fundamental to establishing the amount of effort re-

quired by air traffic controllers to manage and separate aircraft within an airspace. While

taskload measures are unable to account for hidden mental processes in the daily work of

air traffic controllers, the measures do provide a metric by which to understand controller

effort [62].

68
 Taskload can correspond to a number of actions (acknowledgements, hand-offs, clear-

ances, etc); however, the metric studied here is the number of resolution commands issued

to ensure aircraft separation. Consequently, the term conflict-resolution taskload is coined.

Definition 3. Conflict-resolution taskload is the number of resolution commands used

in ensuring separation between aircraft during a given time period.

Already, one study has supported the theoretical framework of addressing conflict-

resolution taskload as a proxy for controller workload when operating under an advisory

conflict-detection and resolution system. For the multi-stage human-in-the-loop study

described in [53], air traffic controllers were tasked with managing a high traffic inten-

sity airspace using current practices and procedures, and later using an advisory conflict-

detection and resolution system. In a second set of experiments, using an advisory conflict-

detection and resolution tool, human simulation studies were performed under a 4D trajec-

tory based operations (TBO) framework with an automated acknowledgement and hand-off

system. Both these operational settings are representative of other proposed concepts within

next generation air traffic systems [50]. Trajectory-based operations is an alternative to the

current clearance-based procedures now found in air traffic control. Instead of aircraft being

cleared to make individual movements along a trajectory (e.g. heading or altitude changes),

complete trajectories are approved by the air traffic controller. In contrast to the current

piecewise planning of clearance-based control, longer-term planning is expected to enable

greater consideration of pilot/carrier preferences and optimal airspace system performance

[50]. Automated acknowledgement and hand-off systems also represent a significant depar-

ture from current practice in which air traffic controllers and pilots are expected to make

vocal contact with one another. By off-loading these tasks to an automated system, hu-

man controllers are able to focus their attention on their primary task of ensuring safety

through conflict-detection and resolution. Studies have already demonstrated the benefits

of automated hand-offs and trajectory-based operations, including significant reductions in

controller workload [12, 13, 53, 68, 79].

A number of key results emerged from the study in [53]. First, when working under

69
 Data Communication

Data Communication

Decision Support Tool Radar

Figure 47: Air traffic controller deferring to an advisory conflict-detection and resolution
system.

current procedures, air traffic controllers are barely able to mange 2X traffic intensities,

and fail at 3X traffic levels. More pertinent to this chapter is that, when operating under

an advisory conflict-detection and resolution system (with TBO and automated hand-offs),

the measure most correlated to controller workload is the number of pending conflicts an

air traffic controller has to resolve, i.e. the conflict-resolution taskload.

Having established that conflict-resolution taskload is worthy of study, it is of value to

note one key statement made by the authors of [53]: “At the 3X level controllers and students

accepted almost all advisories (∼ 98%) due to time pressure.” Effectively, the authors

have indicated that at high traffic levels, air traffic controllers are unable to appropriately

use the advisory system in the manner it was intended. Instead of checking trajectory

solutions for safety, and potentially rejecting and deriving alternatives, the human controller

operates as an open gate, automatically accepting all proposed solutions when overloaded.

This situation is depicted in Figure 47. Implicitly, the advisory human-in-the-loop system

reverts towards an automated system with the conflict-detection and resolution tool directly

managing and separating aircraft. When considering that in the study the conflict-detection

70
and resolution system failed to identify or resolve all potential conflicts, then such deference

behavior is of great concern. As best put by the author of [30], “There is a long history of

cases in which operators are reportedly unaware of automation failures and do not detect

critical system state changes when acting as monitors of automated systems.” Further,

“since the systems to be monitored continue to increase in complexity with the addition of

automation, an increased trend towards large catastrophic failures often accompanies the

incorporation of automation.”

In response to the results above, the research goal of this thesis is to enable the design of

advisory conflict-detection and resolution tools to be more consistent with the capabilities

of human air traffic controllers. By making the best use of information, tools can be

designed to reduce the number of resolution commands required to separate aircraft in

order to manage time pressures. By studying the implementation, capabilities, and policies

behind an advisory conflict-detection and resolution system, the discovery of best practices

is possible.

To better understand the relationship between advisory systems and conflict-resolution

taskload, a three step process is taken. First, in this chapter, graph based models are

introduced and developed to describe the conflict-detection and resolution process. The

graph models are an extension of the work introduced in [41]. In later chapters, human-

in-the-loop conflict-detection and resolution decision-support tools are abstracted to model

their behavior and characteristics. Finally, the abstracted models are applied to the pre-

viously generated traffic scenarios to understand how their implementations can reduce to

conflict-resolution taskload.

4.1 Assumptions

To support the modeling of the conflict-detection and resolution process utilizing graphs,

the following assumptions are required:

• The conflict detection and resolution process can be modeled according to a discrete-

time system with a fixed step-size.

• Conflict relationships are deterministic.

71
 • A dynamically feasible resolution command always exists to resolve potential conflicts.

• Resolution commands are implemented exactly as required and without delay.

4.2 Representing aircraft and potential conflicts through graphs

Consider a set of M aircraft, A = {A1 ...AM }, traveling through an airspace, as illustrated

in Figure 48a. Aircraft trajectories are assumed to occur in 3D space across multiple flight-

levels. According to aircraft intents, each aircraft has the potential to be in multiple conflicts

if no control action is taken. For the en route environment, two aircraft are declared to be

in conflict if they come within 5 NM laterally and 1,000ft vertically of each other. Federal

Aviation Administration regulations require that air traffic controllers issue resolution com-

mands to resolve any identified potential conflicts before they are realized. In the context

of this thesis, a potential conflict is defined according to the definition below.

Definition 4. A potential conflict between two aircraft is defined to exist if, according to

the best available trajectory information and predictions, the two aircraft might come into

conflict if no action is taken by the air traffic controller.

When and how a potential conflict is identified is addressed later in Section 5.3.

Aircraft and potential aircraft conflict relationships are represented by means of a graph.

A representation of the potential conflicts for the example in Figure 48a is given by the

undirected graph, G = (V, E), depicted in Figure 48b. Aircraft are represented by nodes

in the vertex set V = {n1 , ..., nM }, where node ni corresponds to aircraft Ai . Following

regulations, for any pair of aircraft, (Ai , Aj ), that are in potential conflict, the air traffic

controller must issue a resolution command to at least one of the aircraft. Potential conflicts

are indicated by an undirected edge in the edge set E. That is, for the potential conflict

between aircraft Ai and Aj , (ni , nj ) ∈ E. Adhering to the definitions above, the following

definition for conflict graphs is provided.

Definition 5. A conflict graph, G = (V, E), is an abstract representation of aircraft and

potential conflict relationships. Aircraft are represented by nodes in the vertex set V, and

potential conflicts are represented by edges in the edge set E.

72
 x

n3
A3

A1
n1

Track A2
Potential Conflict A4 n2 n4
(a) Aircraft traversing airspace (b) Conflict graph representation

x
n3

A3
n1

A1 n2 n4

Track A2
Potential Conflict A4 Issue resolution command

(c) Proposed solution (d) Proposed solution in the graph

space

n3
A3

A1
n1

Track A2
Potential Conflict A4 n2 n4
(e) Proposed solution implemented (f) Conflict graph updated according to
new trajectory

Figure 48: Example graph representation of aircraft and potential conflict relationships.

73
 Working in this framework, if the graph is completely disconnected, then the airspace

is conflict-free and requires no action from the air traffic controller. Otherwise, edges exist

within the graph and the air traffic controller should issue resolution commands to resolve

potential conflicts. The act of issuing resolution commands to aircraft, assuming appropri-

ate selection and implementation of the maneuvers, removes edges in the conflict graph.

Through a series of controller actions, potential conflicts are resolved and correspondingly,

edges in the conflict graph are removed.

Assertion 1. An airspace is conflict-free if and only if the corresponding conflict graph is

completely disconnected.

A portion of the conflict-resolution process is illustrated in Figure 48 for the example

provided. First, an aircraft is selected to be maneuvered, and a resolution command is

proposed, as indicated in Figure 48c and Figure 48d. Following issuance of the resolution

commands, potential conflicts are considered resolved and edges are removed from the

graph G, as depicted in Figure 48e and Figure 48f. If aircraft A3 is also issued a resolution

command, then it is possible to resolve all potential conflicts, such that the associated graph

is completely disconnected.

The graph representation above is static; however, it can be extended to the dynamic

case. A dynamic conflict graph is able to account for changes in the airspace as aircraft

enter or exit and as potential conflicts are detected and resolved. Ultimately, a dynamic

conflict graph represents the ongoing conflict-detection and conflict-resolution process.

To facilitate modeling of dynamic conflict graphs, a series of terms are introduced. Each

term aids in describing if an aircraft belongs to the vertex set or is part of an element in

the edge set. Furthermore, a distinction is made to indicate if an aircraft is able to accept

a resolution command to prevent conflict.

The dynamic conflict graph is indicated by the discrete-time system, G(k) = (V(k), E(k)).

The set V(k) corresponds to the set of nodes belonging to the graph at time-step k. Simi-

larly, E(k) contains the set of edges.

To construct dynamic conflict graphs, the definitions below are necessary. In accordance

74
with the purpose of the thesis, relevant terms are defined given the existence of an advisory

decision-support tool to help automate the conflict-detection and resolution process.

Definition 6. An aircraft is visible when the decision-support tool is aware of the aircraft’s

presence within or nearby the relevant airspace. If an aircraft Ai is visible at time-step k,

then ni ∈ V(k).

Definition 7. An aircraft is controllable if the air traffic controller is allowed to issue

the aircraft a resolution command. In this case, the decision-support tool is permitted to

generate a resolution command for the aircraft and propose a solution trajectory to the air

traffic controller. According to standard operating procedures, an aircraft is controllable

when it is located within the airspace of the managing air traffic controller or handed-off by

the prior controller of the adjacent sector.

Definition 8. A potential conflict between two aircraft, (Ai , Aj ), is visible if the decision-

support tool predicts and identifies the potential for loss of separation. In this case, if a

potential conflict between Ai and Aj is visible at time-step k, then (ni , nj ) ∈ E(k).

Definition 9. A potential conflict between two aircraft, (Ai , Aj ), is resolvable at time-step

k if either Ai or Aj is controllable for the time-step.

In the dynamic case, the conflict graph grows and shrinks incrementally with the arrival

and departure of aircraft, and the identification and resolution of potential conflicts. Begin-

ning with the conflict graph G(k) = (V(k), E(k)), the graph is updated with the occurrence

of each event (e.g. arrival, departure, identification of a potential conflict) during each time-

step. For the time-step k, V + (k) corresponds to the set of new aircraft arriving into the

airspace over the considered interval of time. Any new potential conflicts identified within

the same time-step are described by the edge set E + (k). The new edges in E + (k) include

both potential conflicts generated by the arriving aircraft in V + (k), and any new conflicts

detected from aircraft already in the airspace. As such, E + (k) may contain nodes (aircraft)

that were already in the graph (airspace) prior to time-step k. During the conflict-detection

process, it is also possible that updated information and trajectory predictions indicate that

75
a potential conflict will no longer occur. The set of unrealizable potential conflicts is desig-

nated by E − (k). Following similar notation, V − (k) denotes the set of nodes corresponding

to the aircraft exiting the airspace.

For the dynamic graph model, resolution commands are not issued at the beginning

of time-step k, but rather throughout it. Thus, the resolution process during time-step k

is not considered completed until the end of the interval. Consider a temporary graph,

G T (k) = (V T (k), E T (k)), just prior to time-step k + 1, representing the airspace before any

resolution commands have been issued, but taking into account other events within the

airspace. In this case, the temporary graph is given by

V T (k) = (V (k) ∪ V + (k))

(15)
T + −
E (k) = (E(k) ∪ E (k)) \ E (k).

The temporary vertex set V T (k) contains aircraft already in the airspace, and new aircraft

entering the airspace. The edge-set E T (k) is representative of existing unresolved potential

conflicts and new potential conflicts that have been identified.

Suppose that conflict-resolution actions act on the temporary graph G T (k). Let CRP

represent an arbitrary conflict-resolution procedure (CRP). The procedure takes as part of

its input G T (k) and returns M, the set of nodes corresponding to maneuvering aircraft.

The set M is dictated by the underlying conflict-resolution policy such that

M = CRP (G T (k)). (16)

Based on the conflict-resolution procedure, the edge set is updated to provide the graph

G(k + 1) at the next time-step. For a resolution command issued to aircraft Ai to prevent

a conflict with Aj , where ni ∈ M, assume that the new advisory trajectory also prevents

future conflicts according to the information encoded in the graph G T (k). So if Ai is

issued a resolution command, then the edge (ni , nj ) is removed from the temporary graph.

Additionally, it is possible that other edges that satisfy (ni , ∗) ∈ E(k) are removed as

well, depending on the resolution command selected. Accordingly, let E r (k) represent any

removed edges by means of the conflict-resolution procedure such that G(k + 1); the graph

at the next time-step is defined by

76
 V(k + 1) = V T (k) \ V − (k)
(17)
T r
E(k + 1) = E (k) \ E (k).

The process described above does not require that all potential conflicts detected by

the advisory system be resolved at the current time-step. But, at the very least, pending

potential conflicts that will become realized in the next time-step without action must be

resolved to maintain a safe airspace.

Note that the modeling here makes no distinction between trajectory-based operations

and clearance-based control of aircraft. While the modeling procedure accurately accounts

for the number of commands issued to resolve potential conflicts, in the case of clearance-

based control, any subsequent maneuvers to place the aircraft back on its intended path

are not accounted for when using lateral maneuvers (e.g. vectoring/heading commands). A

more detailed discussion on the topic is provided in Section 5.7.

Ultimately, by following the discrete-time dynamic model dictated by the graph G(k),

modeling the conflict-detection and resolution process is possible.

4.3 Review of Conflict Graphs

The need for decision-support tools to aid air traffic controllers in the conflict-detection and

resolution process has motivated many researchers to create systems that provide advisory

support. While they transform controller practice and reduce controller workload, there is

limited understanding of how best to design and implement advisory decision-support tools.

The design of an advisory system is particularly important when one considers that no

guaranteed safe conflict-resolution algorithms exist to separate aircraft within an airspace.

Therefore, before designing a conflict-resolution algorithm, it is important to understand

the workload implications of any such system. This chapter takes a first step towards

understanding the relationship between advisory tools and conflict-resolution taskload by

modeling the conflict-detection and conflict-resolution process using graphs. In the ensuing

chapters, the graph modeling is exploited to assess the amount of effort required by air

traffic controllers using a given decision-support tool to resolve potential conflicts.

77
 While modeling potential conflicts through graphs is based on fundamental relationships,

the temporal dynamics rely on a number of assumptions and process constraints. Because

the dynamic graphs are discrete-time systems, the process by which aircraft arrive and

depart and potential conflicts are identified and resolved within each time-step are grouped

together. As a result of the modeling, the order in which each task is processed within

a time-step is not defined, however, some basic ordering is implicit. In the case when

an aircraft enters an airspace, the aircraft should be visible to the conflict-detection and

resolution system prior to being controllable - otherwise, conflict-resolution operations on a

non-visible aircraft is nonsensical. Regardless, the task ordering is considered to be an aspect

of the implemented conflict-detection and resolution program and thus, is left as general as

possible. Furthermore, while not directly addressed in this chapter, future modeling and

implementation of the conflict-detection and resolution processes on the graphs assume a

constant time-step between subsequent values of k. While most research in the current

literature on automated conflict-detection and conflict-resolution tools make use of fixed

time-steps or event-based modeling [16, 83, 100, 101], there are no requirements for such

implementations.

78
 CHAPTER V

ABSTRACTION OF CONFLICT-DETECTION AND RESOLUTION

TOOLS

In the previous chapter, a dynamic graph model, G(k) = (V(k), E(k)), representing air-

craft and their potential conflict relationships was introduced: aircraft appear as nodes,

and potential conflicts are indicated by edges. At each time-step, the vertex and edge sets

are updated as aircraft enter and exit the airspace, and potential conflicts are identified

and resolved. Additionally, a generic conflict-resolution procedure (CRP ) was introduced

to act on the conflict graph as an advisory controller. There is a remaining need to de-

tail how potential conflicts are identified, and how conflict-resolution algorithms resolve

potential conflicts and prevent future secondary conflicts. Instead of specifying the exact

implementation of an advisory conflict-detection and resolution tool, this chapter focuses

on developing abstractions of the behaviors and capabilities of the tool acting on the dy-

namic conflict graphs. Through proper abstraction and parameterization, conflict-resolution

taskload properties for many types of advisory systems can be explored. In this chapter, a

framework for abstracting and parameterizing conflict-detection and resolution systems is

presented.

Abstracting the behaviors and characteristics of conflict-detection and resolution tools

enables high-level assertions on their performance. Ideally, this abstraction of advisory

decision-support tools serves as an alternative to exhaustively designing tools, implement-

ing them in high-fidelity simulations, and analyzing their conflict-resolution taskload. Such

an approach of simulating specific conflict-detection and resolution systems limits the type

of conclusions that can be drawn concerning the design of more generic algorithms: any

results concerning conflict-resolution taskload are predicated on a specific implementation.

By abstracting the properties of conflict-detection and resolution decision-support tools,

79
and implementing simulations in a framework consistent with their abstracted representa-

tions, broader statements on the relationship between the design of algorithms and conflict-

resolution taskload, as well as best practices, can be investigated. Based on the discovery

of best practices, a more guided approach can then be taken in the design of future conflict-

detection and resolution systems.

In this thesis, conflict-detection and resolution systems are abstracted and parameterized

according to the following properties:

1. The aircraft spacing distance used to identify potential conflicts.

2. The growth rate of uncertainty in predicting aircraft trajectories.

3. The policy a conflict-resolution algorithm uses in deciding which aircraft to maneuver.

4. The amount of future trajectory information used in making conflict-resolution deci-

sions.

5. The length of time resolution trajectories can be guaranteed to be conflict-free.

6. The solve-time of the conflict-resolution algorithm. And by extension, the rate at

which trajectory information is updated in the conflict-resolution process.

Properties 1 and 2 relate strictly to the conflict-detection process. Specifically, they cor-

respond to the capability of position sensing systems and trajectory prediction tools used

to measure and forecast aircraft positions. Based on the capabilities and limitations of

trajectory prediction and measurement systems, policies and regulations that set aircraft

spacing distances are established. The next three properties are primarily associated with

conflict-resolution algorithms; however, they also depend on trajectory prediction tools.

Furthermore, while the policy and information used to make decisions is a design property

with some leeway, the amount of conflict-free time a resolution system can guarantee is an

active area of research strongly correlated to the relative “strength” of an algorithm. Finally,

the solve-time is a coupling factor between the conflict-detection and conflict-resolution pro-

cesses.

80
 It is worth noting that the assessment of conflict-resolution taskload (i.e. the number

of resolution commands used to separate aircraft) for a conflict-resolution tool cannot be

made independent of a conflict-detection system. Conflict-resolution decision-support tools

require information concerning which aircraft are in the airspace, what are their intents (i.e.

flight-plans), where potential conflicts are located, and which conflicts require resolution im-

mediately. Because conflict-resolution tools do not have access to this information natively,

they must rely on conflict-detection probes for information when generating solution trajec-

tories. Thus, when analyzing advisory decision-support tools for conflict-resolution, both

the detection process and the resolution process must be considered concurrently.

Further, for conflict-detection and resolution systems, the manner in which potential

conflicts are identified is critical. Minimally, a conflict-resolution algorithm should be ca-

pable of resolving near-term pending potential conflicts that are passed to the algorithm

as constraints by the conflict-detection system. Potential conflicts that are too far into the

future and contain too much uncertainty can be temporarily ignored until more accurate

information is available. Otherwise, if all identified potential conflicts are immediately for-

warded to a conflict-resolution tool and required to be resolved, then the number of advisory

resolution commands can result in superfluous conflict-resolution taskload. That is, poor

handling of uncertainties in aircraft trajectories can identify false potential conflicts that

never arise.

For the remainder of the chapter, the abstraction of conflict-detection and resolution

systems is made in regards to the properties listed above. In the next sections, a receding-

horizon control model, consistent with the dynamic conflict graph representation, is pro-

posed for describing the process by which the conflict-detection and resolution occurs; a

framework for abstracting how potential conflicts are identified is described; and finally,

the capabilities of conflict-resolution systems are then parameterized. In abstracting the

implementation, capabilities, and strategies of the conflict-detection and conflict-resolution

systems, the advisory decision-support tools can be parameterized according to how they

identify and resolve potential conflicts.

81
5.1 Assumptions

Modeling the abstraction of conflict-detection and resolution tools requires a number of

assumptions to enable consistent comparison between implementations. Moreover, mini-

mum standards for behaviors are required. The assumptions supporting the abstraction of

advisory conflict-detection and resolution tools are as follows:

• The conflict-detection and resolution process, as well as the actual implementation of

a conflict-detection and resolution advisory system, can be modeled according to a

discrete-time system with a fixed step size operating according to a receding-horizon

control framework.

• Potential conflict relationships between aircraft are deterministic.

• Potential conflicts, once labeled as not requiring resolution due to updated position

information and trajectory predictions of aircraft, do not reappear as requiring reso-

lution later in time.

• Uncertainty in forecasting aircraft positions grows isotropically at a constant rate in

the horizontal plane. There is no uncertainty in the prediction of aircraft positions

in the vertical direction. Based on the uncertainty model, a cylindrical safety region

around aircraft grows radially.

• A dynamically feasible conflict-free trajectory of predetermined time length exists and

is generated by the conflict-resolution tool for each potential conflict. Furthermore,

aircraft resolution commands ensure that potential conflicts are cleared, and not just

avoided (see Section 5.7).

• Uncertainty in trajectory prediction does not affect the ability of conflict-resolution

tools in generating conflict-free trajectory solutions.

• The conflict-resolution algorithm has the most recent trajectory prediction informa-

tion when it begins resolving any potential conflicts.

82
 • The advisory resolution trajectories are considered to be acceptable to the air traffic

controller and pilot, because resolution commands are guaranteed conflict-free for a

finite time period,

• The air traffic controller implements resolution commands as suggested by the advi-

sory decision-support tool.

• Advisory resolution commands are communicated and implemented exactly as re-

quired and without delay by the air traffic controller and pilot.

The last three assumptions concerning the implementation of advisory commands re-

quire a brief discussion. Despite assuming the air traffic controller implements each advisory

resolution solution as proposed by the conflict-resolution tool, that does not imply that the

air traffic controller lacks the time to verify the safety of each resolution command. Fur-

thermore, the case of rejecting advisory resolution commands is currently ignored. As one

assumption states, advisory resolution commands are assumed to be acceptable to the air

traffic controller. Such a property is only possible through a well-designed conflict-detection

and resolution system.

5.2 Receding-Horizon Control

For the dynamic conflict-graph model presented in Section 4.2, the process of iteratively

identifying and resolving potential conflicts corresponds to a receding-horizon control frame-

work. As used in this thesis, Figure 49 illustrates the corresponding implementation schematic

for advisory conflict-detection and resolution tools using receding-horizon control.

Starting at the current time t, a conflict-detection tool looks at least TR minutes into

the future to check for any potential conflicts. This advanced planning is to ensure that

if resolution commands are issued, then they remain dynamically feasible and safe, even

if there is a delay in implementation by the pilot. Additionally, the buffer time provides

the air traffic controller with sufficient time to generate an alternative resolution command

if he or she elects to reject the advisory commands. The selection of TR allows for some

relaxation of the last three assumptions in Section 5.1.

83
 TR
δt
HD
HR

t Time

Decision-Horizon Time (HD )

Conflict-Free Resolution Time (HR )
Implemented (δt)
Figure 49: Conflict-resolution problem solved in a receding-horizon control framework.

When potential conflicts are detected, the conflict-resolution tool can limit its scope

to only consider a subset of potential conflicts. Specifically, those pending conflicts that

occur within a finite time window of length HD . Limiting the scope prevents unnecessary

consideration of uncertain potential conflicts that are too far into the future. Furthermore,

reducing the scope enables real-time computation of resolution commands for a small set of

aircrat and potential conflicts. Based on the potential conflicts within the HD time window,

the conflict-resolution problem is solved. If HD is too large, then the conflict-resolution

problem can be computationally difficult to solve.

According to the conflict-resolution algorithm, a set of advisory solution trajectories are

generated. The modeling procedure assumes that each trajectory solution is guaranteed

to be conflict-free for at least HR minutes beyond the initial TR if implemented by the

controller.

In the case of receding-horizon control, only the resolution commands associated with

the potential conflicts occurring within a δt time-window are implemented. Resolution of

potential conflicts beyond the [t + TR , t + TR + δt] implementation window are postponed

until future time-steps when more accurate trajectory information is available.

The complete conflict-detection and resolution process is then repeated δt minutes later,

ad infinitum. At each δt time step, the positions of aircraft are updated. This ensures that

up-to-date information on potential conflicts and aircraft trajectories is always available,

thereby preventing any unnecessary resolution commands due to uncertainty. The meaning

84
 min
Dsep /2

1000ft

Aircraft position

Figure 50: Safety region around an aircraft (not to scale)

and implementation behind the parameters HD , HR and δt are later explored in Section 5.3

and Section 5.4 in relation to the abstraction of conflict-detection and resolution tools.

As a requirement for the receding horizon-control framework, HR ≥ HD ≥ δt. Other-

wise, potential conflicts are not identified with sufficient time to resolve them before they

become realized. Also, because the implemented resolution commands are guaranteed to

exist for at least TR + HR minutes, potential conflicts cannot occur between current time t

and t + TR .

5.3 Parameterization of Conflict-Detection Systems

Before applying any conflict-resolution algorithms, potential conflicts must first be detected.

The conflict-detection process is abstracted according to the following properties:

• How potential conflicts are identified.

• How uncertainty affects trajectory predictions.

min and D r
Based on these characteristics, two parameters Dsep sep are identified to describe

conflict-detection tools.

Aircraft spacing requirements can be thought of as placing a cylindrical safety region

around each aircraft that no other aircraft’s safety region is permitted to penetrate. For an

aircraft’s current position, Figure 50 depicts the safety region surrounding it. The radius
min , while the
of the safety region is set by the lateral aircraft spacing requirement of Dsep

vertical buffer is set to 500 ft in each direction. Safety regions are used to classify if

potential conflicts exists. When trajectory predictions forecast that the safety regions of

two aircraft will overlap at some point in time, the aircraft are said to have a potential

conflict. Figure 51a depicts an example of a potential conflict between two aircraft flying

85
 Projected Position
min r
Dsep + Dsep ∆τ

min
Dsep
Current Position Current Position
(a) Potential conflicts are indicated by (b) Trajectory prediction with uncer-
overlapping safety regions tainty

Figure 51: Parameters determining labeling of potential conflicts.

at the same flight-level. To consider uncertainty in trajectory-prediction algorithms, the

radius of the safety region is modeled to grow linearly in time, as depicted in Figure 51b.
r , the radius of the safety region, when projected ∆τ
For a constant growth rate of Dsep
min at the current time, to D min + D r ∆τ in the future.
minutes in advance, grows from Dsep sep sep

Again, potential conflicts are identified according to overlapping safety regions. Growing

the safety regions over time around the predicted aircraft positions is a robust approach to

handling uncertainty in aircraft positions.

min and D r are closely related to the capabilities of aircraft position sens-
The values Dsep sep

ing systems and trajectory-prediction tools. As stated previously, high altitude en route

regulations state that aircraft should always maintain a minimum separation of 5 NM. In

practice however, air traffic controllers begin to issue resolution commands when it appears

aircraft trajectories will come within 8-9 NM of each other. While 4 NM, and even 2 NM

separations are still far from a mid-air collision, uncertainty in radar systems inhibit air

traffic controllers from spacing aircraft at these tight distances. For common en route radar

systems, position updates arrive approximately every 12 seconds from a filtered radar data

stream [76]. Because radar position reports degrade with distance, over 250 NM, “aircraft

less than 5 miles apart could theoretically swap positions on radar monitors [76].” So while

86
radar systems may indicate clearance at 5 NM, noises in sensor measurements can obscure

the fact that aircraft are much closer, even if filtering algorithms (e.g. alpha-beta filters) are

present to estimate aircraft positions [102]. Ultimately, future reductions in aircraft spacing

requirements will require a safety assessment that considers radar error, as well as other

factors such as potential stresses and failures (airspace structure, controller/aircrew work-

load, communication systems), in order to limit the acceptable risk for a mid-air collision

[60]. Further, air traffic controllers add their own safety buffers to ensure they adhere to

safety regulations. In designing advisory conflict-detection and resolution tools, the conflict-

detection sub-system will be required to mimic the same behavior as air traffic controllers.
min is initially tied to aircraft separation requirements (i.e. 5 NM), it is best to
While Dsep

think of the term in regards to how conflict-detection and resolution systems space aircraft.
min parameter,
That is, if an advisory system detects a potential conflict according to the Dsep

then it advises a resolution trajectory to the air traffic controller. Note, however, in practice

the air traffic controller is not required to implement the proposed trajectory if they feel it

is unnecessary. This potential case occurs when the conflict-detection and resolution system

predicts a potential conflict, yet the air traffic controller is confident one will not occur.

The growth rate model used to represent uncertainties in trajectory prediction is a

simplification based on prior studies. Recent works have indicated that the along-track

error in trajectory position grows linearly along straight paths [34, 42, 74] - hence, the use
r . The short-coming of such an approach
of a linear model based on the parameter Dsep

is that uncertainty is assumed to grow isotropically in the X-Y plane. Additionally, the

model assumes no uncertainty in the vertical direction. This simplified model does not

replicate the types of errors associated with position reporting systems and trajectory-
min and D r
prediction tools [96, 67, 39]. Nonetheless, the parameters Dsep min enable a basic

model to understand how potential conflicts are labeled according to aircraft spacing and

uncertainty.
min , the aircraft spacing distance indicating advisement
Based on the two parameters Dsep
r
of a resolution command; and Dsep the growth rate in position uncertainty, the behavior
min is the aircraft
of the conflict-detection process can be parameterized. The parameter Dsep

87
spacing distance used to label potential conflicts that require advisories. Based on the

labeling of potential conflicts, the conflict-resolution algorithm then generates resolution

trajectories.

The following definitions are now provided.

min is the aircraft spacing used in defining safety regions

Definition 10. The parameter Dsep

around each aircraft. When the safety regions encircling the forecast positions of the two

aircraft overlap in space-time, the conflict-detection tool dictates that according to current
min
information the conflict-resolution tool advise a resolution command. The parameter Dsep

should be set to a value greater than or equal to the minimum required separation distance.

r
Definition 11. The uncertainty parameter Dmin represents the growth in trajectory pre-

diction errors in the X-Y plane according to a linear model.

5.4 Parameterization of Conflict-Resolution Systems

Much like conflict-detection tools, the capabilities and behaviors of conflict-resolution al-

gorithms can be abstracted and parameterized. Here, the abstraction of conflict-resolution

tools focuses on the following properties:

• How much information is used in decision making.

• The decision-making policy used for selecting which aircraft to maneuver.

• Tool capability for finding conflict-free resolution commands.

• How often the conflict-resolution problem is solved.

The abstraction introduces two additional parameters, HD and HR , henceforth referred

to as the decision-horizon time and the conflict-free resolution time. Additionally, the

abstraction also makes use of δt, the solve-time parameter. The parameters HD , HR , and

δt are first described assuming the presence of a conflict resolution procedure (CRP), such

as the one introduced in Chapter 4. Next, the decision policies forming the backbone of

conflict-resolution procedures are detailed.

88
5.4.1 Implementations and Capabilities: HD , HR , and δt

HD : Decision-Horizon Time. In Chapter 4, a conflict-resolution procedure acting on

a dynamic conflict graph is described according to Equation 16. Taking as its input the

temporary conflict graph at the current time step, G T (k), the conflict-resolution procedure,

CRP , outputs resolution commands to a set of aircraft M. That is, M = CRP (G T (k)). It

is important to note that the decision of which aircraft to maneuver need not be based on

all information supplied to the conflict-resolution procedure. In fact, if a conflict-resolution

procedure did consider all aircraft and potential conflicts, even those far into the future,

there exists the possibility that the conflict-resolution problem would no longer be com-

putationally tractable. For example, when considering simultaneous conflict-resolution in

the horizontal plane, for n aircraft involved in potential conflicts, the number of possible

discrete decisions scales according to n(n − 1)/2. At 2X traffic levels, it is not uncommon

to have up to 40 aircraft in a sector. When considering discrete decision variables, and

any other necessary calculations (fuel-burn, topologically similar yet alternative routes),

the conflict-resolution problem is challenging to solve simultaneously for all 40 aircraft in

implementation-time, even with simple algorithms [100]. Section 5.5 has a more detailed

discussion on the issue of solution times.

To simplify the search for resolution commands, the conflict-resolution procedure can

trim G T (k) to limit the amount of information it considers in generating resolution trajec-

tories. Trimming of temporary conflict graph G T (k) yields a sub-graph referred to as the

decision graph, denoted Ĝ(k). The size of the decision-graph is established by the parameter

HD , the decision-horizon time. Now, instead of considering all potential conflicts, only a

subset of them are considered when solving the conflict-resolution problem. More formal

definitions of the decision-horizon time and decision-graphs are provided below.

Definition 12. The parameter HD , the decision-horizon time, as measured in minutes,

indicates the time range over which conflict-resolution algorithms consider potential conflicts

that might result in multiple aircraft resolutions.

Definition 13. The decision-graph, Ĝ(k) = (V̂(k), Ê(k)), at time-step k, for the current

89
 Track
Conflict considered in decision−making N3

A3

N1
A1
A2 N2 N4
A4
(a) Small value of HD (b) Ĝ(k): Small value of HD

N3

A3

N1
A1
A2 N2 N4
A4
(c) Larger value of HD (d) Ĝ(k): Larger value of HD

Figure 52: Representation of potential conflicts that are considered in the decision-making
process.

time t, and parameterized according to the decision-horizon time HD , is constructed by

V̂(k) = V T (k) (18)

Ê(k) = {(ni , nj )|(ni , nj ) ∈ E T (k) and tci,j ∈ [t, t + HD ]}, (19)

where the values tci,j are the predicted potential conflicts times given by Equation 4.

A better understanding of the use of decision graphs and the decision-horizon time

can be gained through the examples illustrated in Figure 52. When HD is small, the

conflict-resolution procedure only considers a single potential conflict between aircraft A1

and A2 (see Figure 52a and Figure 52b). Consequently, the conflict-resolution procedure

generates a resolution command for one of the aircraft assuming that aircraft A3 and A4

continue on their intended trajectories. When HD is larger, the decision-graph is expanded

to include the potential conflicts between aircraft A1 and A4 , and between aircraft A2 and

A3 . Thus, when deciding how to resolve the potential conflict between A1 and A2 , the

future resolution actions of A3 and A4 are considered, assuming that they are required to

90
maneuver. In contrast, for the prior case with a small value of HD , the conflict-resolution

algorithm assumes aircraft continue on their intended paths.

As noted previously, even the internal calculations of conflict-resolution tools require

interaction with conflict-detection tools beyond the initial labeling of potential conflicts.

Because the value of HD relates to how far in advance potential conflicts are considered,

selecting a value of HD must be within the capabilities of the trajectory-prediction tools;

potential conflicts cannot be considered if they cannot even be identified.

HR : Conflict-Free Resolution Time. The capability of conflict-resolution algorithms

is perhaps most related to the ability to ensure separation between aircraft. If an algorithm

is unable to generate resolution commands that maintain proper spacing, then its value

as an advisory decision-support tool is quite limited. Furthermore, if a conflict-resolution

algorithm poorly routes aircraft, thereby generating secondary conflicts, then the advisory

tool might create more work than it was intended to resolve. In parameterizing the strength

of conflict-resolution algorithms, the guaranteed conflict-free resolution time, HR , is intro-

duced. The value of HR represents the amount of time a conflict-resolution algorithm can

guarantee that a resolution command is conflict-free beyond the initial TR minutes.

The importance of the conflict-free resolution time HR is illustrated in Figure 53. When

resolving a potential conflict, there is little value in issuing a resolution command to pre-

vent one conflict if it generates another. As such, the relative size of HR is important.

Furthermore, HR can be related to issues of safety. If conflict-free resolution times can

only be guaranteed for short time-horizons, then air traffic controllers might be required to

continuously resolve conflicts.

Definition 14. The parameter HR , the conflict-free resolution time, as measured in

minutes, indicates the amount of time conflict-resolution algorithms guarantee resolution

commands to be conflict-free beyond the initial TR minutes.

While the conflict-free time is stated as guaranteed, this is an assumption used only for

91
 A3
A1

Potential Conflict
A2
(a) Original traffic scenario

Resolved Path

A3
A1

Guaranteed
Conflict−Free
A2
(b) Resolved scenario with small HR generating sec-
ondary conflict

Resolved Path

A3
A1

Guaranteed
Conflict−Free
A2
(c) Resolved scenario with larger HR avoiding secondary
conflict

Figure 53: Larger values of the guaranteed conflict-free time, HR , help to prevent secondary
conflicts.

92
mathematical modeling purposes. There are no guaranteed conflict-free algorithms. Fur-

thermore, the lack of provably safe algorithms is one argument supporting the implemen-

tation of advisory systems. Therefore, as an alternative conceptualization, the guaranteed

conflict-free time can be thought of as an expected value that is satisfied with probability

1 − , where  → 0.

δt: Solve-Time. The parameter δt is used to describe how often the conflict-resolution

problem is solved. The solve-time of a conflict-resolution algorithm is relevant because

of how uncertainty in trajectory predictions manifests in excessive labeling of potential

conflicts. As modeled in Section 5.3, uncertainty in trajectory predictions leads to the

identification of potential conflicts which might not occur. While a rapid update rate for

trajectory prediction is helpful in detecting potential conflicts, the conflict-resolution tool

can only make use of information available when it begins solving the conflict-resolution

problem.

Consider the case when the conflict-detection system prompts a conflict-resolution al-

gorithm to resolve a number of potential conflicts. Also suppose that while the conflict-

resolution problem is being solved, the conflict-detection system updates the state of po-

tential conflicts such that one potential conflict is identified to no longer occur. Because

conflict-resolution algorithm is running, the updated information cannot be considered, and

thus, a resolution advisory is generated for the no-longer existent potential conflict.

The conflict-resolution algorithm gets snap-shots of the airspace, and must make de-

cisions using the best available information. An illustration of this idea is provided in

Figure 54. From the perspective of the conflict-resolution tool, the current positions and

trajectories of aircraft are updated every δt minutes. And likewise, the safety region around

each aircraft is reset.

The following definition of δt, the solve-time, is provided.

Definition 15. The parameter δt corresponds to the solve-time, or how often the conflict-

resolution problem is solved. Accordingly, aircraft position and trajectory prediction, and po-

tential conflict information is updated within each implementation of the conflict-resolution

93
 Predicted Position
2δt later
Actual Trajectory
Predicted Trajectory Predicted Position
δt later

Original
Trajectory Prediction

Updated Predicted Position

Trajectory Prediction δt later

Figure 54: With solve-times of δt, trajectory predictions are updated for each implementa-
tion of the conflict-resolution tool.

algorithm.

In Section 5.5 a discussion of solution times and the parameters HD and HR is provided.

5.4.2 Conflict-Resolution Policies

Once potential conflicts are identified, the role of the advisory conflict-resolution algorithm

is to select which aircraft should receive resolution commands. This section details three

underlying policies for selecting aircraft. The first policy implements a simple first-come,

first-served (FCFS) heuristic. The second policy randomly selects aircraft to maneuver until

all potential conflicts are resolved. The final policy, is the Minimum Conflict-Resolution

Taskload (MCRT) policy, which minimizes the number of resolution commands used to

deconflict traffic.

Other conflict-resolution policies are detailed in Appendix A, all of which are ruled-based

heuristics policies like the first-come, first-served and random policies.

The three policies highlighted in this section are specifically chosen to better understand

the bounds on conflict-resolution taskload. The MCRT policy minimizing the number of

resolution command proposed at each time-step. Thereby approaching a lower bound on the

conflict-resolution taskload when implemented over long time periods. The FCFS policy is

94
representative of other generic rule-based heuristic policies. And finally, the random policy

serves as a pseudo-upper bound and baseline standard for determining if other policies result

in excessive conflict-resolution taskload due to their formulation.

Each conflict-resolution procedure, operating in accordance to a specific policy and

within the receding-horizon framework, receives as part of its input a description of the

conflict-relationships identified by the conflict-detection system. Specifically, the input

includes the temporary conflict-graph, G T (k), for the current time-step. Following the

trimming process detailed in Section 5.4.1, parameterized by HD , the conflict-resolution

procedure then forms the decision graph, Ĝ(k) = (V̂(k), Ê(k)). Aircraft that are visible to

the automated system at the time-step k, are represented by the vertex set V̂(k). Potential

conflicts are represented by the edge set Ê(k).

A working definition of conflict-priority ordering is required.

Definition 16. A conflict-resolution policy with conflict-priority ordering resolves po-

tential conflicts according to the order in which they are predicted to occur.

The decision policies presented in this thesis are representative of non-cooperative conflict-

resolution policies that maneuver only aircraft involved in potential conflicts. While there

does not exist a formal definition for cooperative or non-cooperative conflict-resolution poli-

cies, a working explanation is provided here.

Consider a potential conflict between aircraft Ai and Aj that requires at least one

aircraft to maneuver. Let Ci and Cj represent the cost of maneuvering one aircraft or the

other. Additionally, let Ci,j be the cost of maneuvering both aircraft to resolve the potential

conflict. For cooperative conflict-resolution algorithms, the relationship

Ci,j ≤ Ci + Cj (20)

can hold. Accordingly, in the context of an optimization framework, maneuvering both

aircraft is the preferred option over maneuvering a single aircraft. Conversely for non-

cooperative decision policies, it is a necessary condition that

Ci,j < Ci + Cj (21)

95
for each pairwise conflict. Therefore, in the case of a single pairwise conflict, only one

aircraft is maneuvered. When conflict-priority ordering is present, Equation 21 is a sufficient

condition for non-cooperative decision policies.

First-come, first-served. The first-come, first-served policy (FCFS) is a straightforward

heuristic that uses sequential path planning to generate safe aircraft trajectories. When

implemented with conflict-priority ordering, conflicts are resolved sequentially until none

remain. For each conflict-pair, (Ai , Aj ), the aircraft arriving first into the sector is given

priority, and thus, the aircraft arriving later is issued an advisory resolution command.

The algorithm for implementing the FCFS policy with conflict-priority ordering at time-

step k is found in Procedure 2. Procedure 2 is initialized with the decision graph, Ê(k), then

potential conflicts are processed in the order of predicted occurrence. The predicted time of

occurrence for each potential conflict between aircraft Ai and Aj is denoted by the variable

tci,j . Once the nearest pending unresolved potential conflict is identified, the arrival times

for the aircraft are compared to determine which aircraft to maneuver. In Procedure 2,

the temporary decision vector, mT , contains the ordered set of aircraft that are issued

resolution commands according to the FCFS policy with conflict-priority ordering. The

value of mTq refers to the q th aircraft to be issued a resolution command. Because the

receding-horizon control framework only issues resolution commands to those aircraft with

potential conflicts occurring in the [t + TR , t + TR + δt] time interval, later potential conflicts

are left to be solved in the future. The final step of the procedure is to segregate the

aircraft associated with near-term potential conflicts. The output is M, the set of aircraft

the advisory conflict-detection and resolution system will advise rerouting.

96
Procedure 2 First-Come, First-Served with Conflict-Priority Ordering
E T ← Ê(k), q = 0

while card(E T ) do

q =q+1

(i∗ , j ∗ ) = argmin tci,j

i,j

ni∗ if taj∗ < tai∗


T
mq =

n ∗ else

j

ET ← ET \ (mTq , ∗)

end while

I = [t + TR , t + TR + δt]

M = {m | m ∈ mT ∩ ∃tcm,∗ ∈ I}

The removal of the conflict-priority ordering requirement results in a slightly different

process for resolving potential conflicts; the procedure for FCFS without conflict-priority

ordering is provided in Appendix A.

The primary advantages of the FCFS policy is that it provides a straightforward proce-

dure to decide which aircraft are issued resolution commands. In regards to queuing for a

service, the idea of prioritizing aircraft according to entry times appears sensible. However,

one disadvantage of the FCFS policy is that there exists the possibility of consistently favor-

ing aircraft traveling along specific routes. For example, consider the jetroutes illustrated

in Figure 55 for ZMP42. For east-bound traffic along route J64, and north-bound traffic

along route J25, a common point along the jetroutes promotes the possibility of potential

conflicts. When operating under a FCFS policy, in the case of a potential conflict, the air-

craft traveling along J64 are given priority over aircraft on J25 because they arrive into the

sector earlier (as estimated by the distance between the aircraft entry points into the sector

and the potential conflict). Through the FCFS decision mechanism, aircraft flying along

specific routes are consistently favored over others, which could be considered as ‘unfair’

by air traffic carriers that heavily traffic the unfavored routes. While there is no direct

97
 95° W
42° N J25

ZMP43
J64

40° N

Figure 55: Sector map of ZMP42 with jetroutes

statistical evidence supporting persistent bias in automated conflict-resolution, the idea of

fairness is often noted by designers of algorithms [2, 15, 45, 97, 101].

Random Policy. The introduction of a random policy to select maneuvering aircraft

serves two purposes. First, in contrast to the FCFS policy, a random policy for assigning

which aircraft are maneuvered does not contain biases that are favorable to a particular

subset of aircraft. Second, a random policy serves as a reference standard to compare

with the conflict-resolution taskload of other heuristic policies. For example, if the conflict-

resolution taskload of the FCFS policy is greater than that for the random policy, then it

can be stated that the FCFS policy yields excessive conflict-resolution taskload.

98
Procedure 3 Random with Conflict-Priority Ordering
E T ← Ê(k), q = 0

while card(E T ) do

q =q+1

(i∗ , j ∗ ) = argmin tci,j

i,j

ni∗ if rand < .5


T
mq =

n ∗ else

j

ET ← ET \ (mTq , ∗)

end while

I = [t + TR , t + TR + δt]

M = {m | m ∈ mT ∩ ∃tcm,∗ ∈ I}

One possible random policy is described by Procedure 3. The policy randomly selects

which aircraft are issued advisory resolution commands according to conflict-priority order-

ing. Another random policy is proposed in Appendix A.

Minimum Conflict-Resolution Taskload Policy. Thus far, the policies introduced

have relied on fixed procedures to determine which aircraft should be issued a resolution

command. Further, none of the policies explicitly consider the potential repercussions on

controller conflict-resolution taskload. The Minimum Conflict-Resolution Taskload policy

(MCRT), overcomes this short-coming. Furthermore, the policy seeks to established a lower-

bound on the amount of effort required to manage an airspace according to the capabilities

of the conflict-detection and resolution system.

Addressing lower bounds on controller conflict-resolution taskload is fundamental to

establishing the amount of effort required by air traffic controllers to manage and separate

aircraft within an airspace when using an advisory decision-support tool. Currently, the

effort required to resolve conflicts in an airspace remains a relatively ill-defined metric.

Through a graph-based approach, tools from graph theory are applied to determine a lower

99
 x x
n1 n6 n1 n6
n4 n4

n2 n5 n2 n5

n3 n3

ni Removed Node
x x
(a) Initial graph. (b) Resulting graph.

Figure 56: Example application of the minimum vertex cover problem.

bound on the number of resolution commands required to separate aircraft.

Determining the minimum number of resolution commands is equivalent to applying the

minimum vertex cover problem for graphs. According to [104], the minimum vertex cover

problem (MVCP) asks: ‘What is the minimum number of nodes that can be removed from

a graph, such that the remaining graph is completely disconnected?’ The corresponding act

of removing a node ni from the graph G, is equivalent to issuing the aircraft Ai a resolution

command. An example application of the minimum vertex cover problem is shown in

Figure 56. Following removal of the nodes, the remaining graph is completely disconnected

as illustrated in Figure 56b. (Note, the minimum vertex cover problem can have multiple

solutions.) In the case of applying the minimum vertex cover problem to conflict-resolution,

instead of removing nodes from the graph, any edges associated with the node are removed

from the edge set. That is, if aircraft Ai is issued a resolution command, then after applying

the conflict-resolution procedure, (ni , ∗) 6∈ EˆT (k).

The minimum vertex cover problem can be expressed as a Mixed Integer Linear Pro-

gram (MILP). While, there exist efficient large-scale methods for solving the NP-complete

minimum taskload problem [18, 19, 88], given the relative size of the conflict graphs found

in air traffic (i.e. less than 20 aircraft per conflict cluster, see Chapter 3 and [41]), the MILP

formulation can be solved with sufficient speed (i.e. 0.02 seconds) by most standard integer

100
program solvers (e.g. CPLEX or LPSolve, and even Matlab).

For the decision graph Ĝ(k) = (V̂(k), Ê(k)) with M aircraft, let Ri be a binary variable

indicating if aircraft Ai is to be issued a resolution command. If any two aircraft have nodes

within the edge set Ê(k), then at least one of the corresponding binary variables must be

1 to indicate a resolution command is issued to the aircraft. The minimum number of

resolution commands is equivalent to minimizing the sum of the binary variables, Ri . The

corresponding MILP is:

PM
min i=1 Ri

s.t. Ri + Rj ≥ 1 ∀ (ni , nj ) ∈ Ê(k) (22)

Ri ∈ {0, 1} ∀ i = 1 . . . M
The solution to the MVCP for conflict resolution is then given by the indices of the non-zero

elements of the vector R.

The optimization problem in Equation 22 serves as the foundation for the minimum

conflict-resolution taskload policy. The procedure for the MCRT policy is given in Procedure 4.

The function M V CP (Ê(k)) represents the application of the minimum vertex cover prob-

lem. Following the nomenclature previously established, M, is the set of aircraft to be

issued advisory resolution commands during the current time-step.

Procedure 4 Minimum Conflict-Resolution Taskload Policy

mT = M V CP (Ê(k))

I = [t + TR , t + TR + δt]

M = {m | m ∈ mT ∩ ∃tcm,∗ ∈ I}

5.5 Solution Times

As it stands, the parameters HD , HR and δt, have no limitations as formalized. However,

from a practical perspective, the ability to design an algorithm with large values of HD and

HR , and a small value of δt is challenging. That is, considering large amounts of information

and generating long conflict-free trajectories require longer computation times.

101
 Option 1

Aircraft 1 Aircraft 2

Option 2

Figure 57: In the planar case, there exist two topologically distinct options for each resolu-
tion command.

Consider a potential conflict between two aircraft, as illustrated in Figure 57. Assuming

reasonable behavior of aircraft routings (e.g. no loops), there exist two topologically distinct

options for one aircraft maneuver around the other. For the case presented in the figure,

these options are colloquially posed as “left or right.” Recognizing that “left or right” is

a relative term, it does not matter which aircraft is actually issued the maneuver. And

fundamentally, for the planar case there are only two options regardless of the aircraft

configuration (e.g. trailing aircraft: pass or stay behind; crossing aircraft: pass in front

or pass behind) that represent two disconnected solution spaces. When extended to the

simultaneous conflict-resolution problem with an arbitrary set of n aircraft, the number of

combinations for the “left or right” option is given by n(n − 1)/2, i.e. n choose 2.

For optimal formulations, the only way to ensure an optional solution is to exhaus-

tively search through all possible decision-options (e.g. using branch and bound methods

[80]). Thus, as the number of aircraft grows large, the disconnected solution space makes

optimization difficult. This process can be quite time consuming when other options and

constraints are considered. For example, constraints associated with aircraft dynamics re-

strict the size of each disconnected solution space, thereby making it challenging to even

find initial feasible solutions to begin optimizing over (e.g. applying gradient descent meth-

ods). Furthermore, when relaxing which aircraft are allowed to maneuver, the “left or

right” disconnected spaces are further subdivided into at least three options for each air-

craft pair: maneuver aircraft 1, maneuver aircraft 2, maneuver both aircraft, maneuver no

aircraft (assuming this option is allowable). With each disconnected sub-space, solution

times generally increase. In the worst case, solution times grow exponentially.

Rule-based heuristic methods for conflict-resolution attempt to overcome solution-time

102
issues associated with optimal formulations by providing a fixed search procedure through

a library of possible trajectory solutions. Because of the fixed library size, and a prescribed

and dictated heuristic search procedure, rule-based methods often run in real-time [32, 73].

However, the capability of rule-based algorithms in finding feasible solutions is directly

related to the heuristic search procedure and the size of the solution library. Furthermore,

because of the difficulty of encoding descriptions of multi-aircraft conflicts into a rule-

based search, these methods only consider pairwise conflicts when generating resolution

commands. Thus, when resolving one potential conflict, possible secondary effects related

to other potential conflicts are not considered.

There are key distinctions between rule-based and optimal formulations in their solu-

tion times and implementations. The rule-based methods require that HD = 0 minutes, as

only individual potential conflicts are considered at each time-step. Optimal formulations

allow HD to be set at any desired value. However, as HD grows, the number of potential

conflicts considered in the optimal formulation increases. As such, the solution space be-

comes increasingly disconnected (scaled by n(n − 1)/2), thereby making it difficult to solve

in real-time. So, from a practical perspective, increasing HD is not always feasible. Of

course attempts around this problem have been made by a number of researchers, partic-

ularly through the application of evolutionary optimization algorithms to rapidly explore

the disconnected solution space [28].

In the degenerate case when HD = 0 for the rule-based heuristics, a set of n problems

is solved with 2(n − 1) binary decision-options. While requiring a greater number of binary

decisions than the simultaneous problem (n(n − 1)/2 total binary decisions), the complete

set of smaller problems can be solved faster than the simultaneous formulation [4].

Increasing the value of HR is similarly constrained by the requirement for advisory tools

to issue resolution commands in real-time. Consider the dynamic routing problem illus-

trated in Figure 58, in which an aircraft must find a path through moving weather (moving

weather is a proxy for other aircraft). At each point along the aircraft trajectory, x(t),

the constraint x(t) 6∈ W(t) ∀t (where W(t) represents weather blockages) must be satis-

fied at each time-step. For the path planning problem, as HR increases, the number of

103
 x(t)

Figure 58: Finding solution to the motion planning problem becomes increasingly difficult
as time horizons increase.

weather blockage constraints scales linearly. Meanwhile, additional constraints that ensure

trajectory solutions satisfy aircraft dynamics (e.g. turn rates and stall speeds) are included.

Furthermore, what makes the path planning problem even more difficult is that trajecto-

ries around moving weather (or aircraft) are inherently nonlinear, nonconvex, and multi-

segmented. These properties prevent application of fast linear optimization algorithms (e.g.

simplex) in solving the routing problem, and they require any rule-based algorithm to have

a large library of solutions that include multi-segment routing.

In summary, large values of HR and HD are preferred for improved optimality and

longer conflict-free solutions, but computation limitations prevent their implementation in

practice. In many cases, with current computational constraints, the solution times for

formulations with large values of HR and HD cannot be generated in real-time, or at least

within the δt solve-rate.

5.6 Special Cases

r , and D min correspond to

Certain sets of implementation parameters δt, HD , HR , Dsep sep

special cases of interest. These special cases are representative of absolute lower bounds, or

common implementation frameworks proposed by previous researchers.

r
When HD = HR = ∞, and Dsep = 0, the parameterization represents a system with

perfect knowledge and capabilities: all potential conflicts are correctly identified far into

the future without the need for conservatively safe buffers, and all resolution commands are

guaranteed to be conflict-free indefinitely. When the minimum conflict-resolution taskload

104
policy is applied with these parameters, the solution represents the minimum number of res-

olution commands used to separate aircraft. The solution serves as an absolute lower bound

for all policies for a given traffic scenario. Thus, based on the solution to this special imple-

mentation, all other implementations in the remaining parameter space and policy space,

can be compared against a reference standard. Additionally, when δt = 0, the implemen-

tation approximates an event-based conflict-detection and resolution tool: when potential

conflicts are identified, they are resolved immediately without delay in computation.

Assigning HD = 0 and δt = 0 is an implementation of interest that approximates

pairwise sequential conflict resolution common to rule-based heuristic policies. The conflict-

resolution process is considered to be pairwise, because as HD → 0, the decision-graph only

contains a single potential conflict (assuming simultaneity of potential conflicts does not

exist). To ensure feasibility of the implementation, δt → 0 at an equal or greater rate

than HD ; otherwise, within each time-interval not all conflicts are detected. Taking this

case to an extreme, when the guaranteed conflict-free resolution time, HR , is small (e.g.

1 minute), the system behaves reactively ,solving conflicts as they appear and only in the

short-term. Such behavior is similar to the Traffic Collision Avoidance System (TCAS),

which is designed to resolve any last-second conflicts that were not identified and resolved

by the air traffic controller. TCAS represents the last line of safety, and its use is considered

to be a failure. However, a difference between the previous policies described and TCAS,

is that TCAS requires both aircraft to maneuver.

5.7 Clearance-Based Control versus Trajectory-Based Operations

At this stage, with the exception of the specification of conflict-free times, no assumptions

are made concerning the nature of the resolution commands provided by the decision-

support tool. That is, each resolution command can consist of any number and type of

maneuvers (heading, speed, or altitude changes) of arbitrary magnitude. Moving forward,

a distinction is needed between a resolution command and a resolution maneuver.

As an example, consider a potential conflict between two aircraft flying at level flight.

Two possible resolution solutions include a flight level change and a lateral change for one of

105
 Top View
Side View

Original Trajectory
Resolved Trajectory
(a) Vertical resolution solu- (b) Lateral resolution solution.
tion.

Figure 59: Two example resolution solutions for a potential conflict

the aircraft. As illustrated in Figure 59, in the latter case, a lateral resolution command can

consist of multiple heading changes. Lateral maneuvers contain multiple heading resolutions

to ensure that the aircraft returns to its original path. This type of path is refer to as a

‘closed loop’ trajectory. In other cases, if the heading change aligns with a future point in

the flight plan, then a direct routing can be issued. This bypasses segments of the original

flight plan. Because of the distinction between a resolution command and a resolution

maneuver, the discussion of conflict-resolution taskload requires care.

Under current clearance-based control operations, the air traffic controller issues vocal

clearances for each resolution maneuver by an aircraft. So, the conflict-detection and resolu-

tion model proposed in this thesis may account for each resolution maneuver used to prevent

a conflict, but future maneuvers to ensure a close-loop trajectory are not considered. As

such, the number of vocal communications associated with routing is under-represented by

the conflict-resolution taskload. Thus, the conflict-resolution taskload represents a lower-

bound. When only vertical maneuvers are used to deconflict traffic, the number of maneu-

vers issued by the controller approaches the conflict-resolution taskload.

Under the future NextGen system, the manner in which aircraft are issued resolution

commands is likely to change. A major component of the NextGen system is the inclusion

of 4D Trajectory-Based Operations (TBO) as an alternative to clearance-based control.

Instead of approving individual resolution maneuvers, particularly in the case of potential

conflicts, a complete trajectory is approved and communicated to the aircraft. The trans-

mission of the new trajectory information can be communicated in multiple ways; however,

106
 Top View Top View

Desired Desired
Exit Exit
(a) Closed-loop resolution (b) Avoidance maneuver.
command.

Figure 60: Difference between avoidance maneuvers and closed-loop resolution commands.

the most likely manner will be through digital data channels. Adherence to the amended

trajectory is made possible through improved navigation aids such as GPS. Assuming air-

craft conform to their trajectories, then air traffic controllers are only required to track

aircraft that have potential conflicts and to provide additional services (e.g. congestion

and weather alerts) to aircraft as requested. For trajectory-based operations, the conflict-

resolution taskload of each policy more closely approximates the amount of effort required

by the controller. Only in the case of amendments to a trajectory, is the air traffic controller

required to verify it.

The use of closed-loop trajectory solutions is important to many conflict-resolution algo-

rithms. As previously stated in Section 5.1, the model here assumes existence of resolution

trajectories that clear potential conflicts, not just avoid them, as there is a critical distinction

between the two cases. A visual representation that distinguishes between the avoidance

maneuvers and resolution commands (particularly closed-loop trajectories), is illustrated

in Figure 60. For the closed-loop resolution solution in Figure 60a, the aircraft reaches its

destination. In contrast, for avoidance maneuvers there is no guarantee that aircraft are

guided to their desired destinations. Instead, by constantly avoiding conflict, the aircraft is

moved further away from its destination, as illustrated in Figure 60b.

5.8 Review

In a major literature review by Kuchar and Yang [56], the authors created a classification

system to describe conflict-detection and resolution systems. The taxonomy they utilized

107
considered whether the algorithms were probabilistic, deterministic, or robust; the types

of resolution maneuvers allowed (vertical, heading, speed); the solution methodology (pre-

scribed, optimized, force field, manual); and whether the formulations were pairwise or

global. In classifying algorithms, some properties and concerns of the conflict-detection and

resolution tools can be readily identified and compared (e.g. deterministic prescribed algo-

rithms fail to capture uncertainty, however they can be solved rapidly). In a similar fashion,

the abstraction and parameterization of conflict-detection and resolution tools described in

this chapter attempt to capture the capabilities of the algorithms. In doing so, assessments

of the conflict-resolution taskload associated with each implementation are possible.

Building upon the conflict-graph models described in Chapter 4, a framework is de-

veloped to describe how conflict-detection and resolution algorithms operate in real-world

implementation. First, a receding-horizon framework is proposed, in which the conflict-

detection and resolution problem is continuously solved. At each time-step, potential con-

flicts are identified and resolved. The resolution commands for potential conflicts occurring

in the near future are immediately advised to the air traffic controller by the decision-support

tool. Potential conflicts that occur later in the future are still identified and resolved in the

context of immediately pending conflicts; however, advisory commands are not proposed to

the controller. As trajectory information is updated at each time-step, the conflict-detection

and resolution process is repeated.

Following a description of the receding-horizon framework, parameters describing the

behaviors and capabilities of the conflict-detection and resolution systems are provided.

Conflict-detection systems are described according to how potential conflicts are labeled as

requiring resolution. Conflict-resolution systems are parameterized according to how far

in advance potential conflicts are considered and for how long resolution commands are

guaranteed to be conflict-free.

Without actually delineating how the resolution maneuvers are selected, the framework

and parameterization of the conflict-detection and resolution tools has been extended be-

yond that conceived by Kuchar and Yang. By characterizing advisory tools according to key

model parameters and policies, a common methodology for analyzing the conflict-resolution

108
taskload of algorithms is now possible.

109
 CHAPTER VI

SIMULATION MODELING FOR ABSTRACTED TOOLS

In the previous chapters, a model for generating uncontrolled traffic scenarios is proposed, a

graph-based representation of the conflict-detection and resolution process is described, and

finally an abstraction of conflict-detection and resolution tools is provided. To understand

the relationship between conflict-resolution taskload and the design and implementation of

conflict-detection and resolution systems, an appropriate simulation environment is required

to extract relevant information. In this chapter, the simulation model used to implement

the abstracted conflict-detection and resolution tools is described. Based on simulations,

the conflict-resolution taskload for each traffic scenario and implementation is recorded, and

a sensitivity analysis is performed to clarify best practices.

The primary difficulty in constructing a reasonable simulation environment, consistent

with the prior abstractions, is in representing the conflict-detection and resolution pro-

cess without specifying the exact trajectory solutions used in resolving potential conflicts.

When resolution trajectories are not specified, the manner by which secondary conflicts

are generated is ill-defined. Prior to describing the process for simulating the abstracted

conflict-detection and resolution tools, some additional properties related to the conflict-

resolution process must be stated such that approximations can be placed into context.

When an air traffic controller or conflict-resolution tool issues a command, ideally, the

resulting aircraft trajectory reduces or removes the possibility of future potential conflicts
o (A |tc ∈ [t, t + ∆τ ]) denote the probability for aircraft A to
in the near-term. Let Pcon i i,∗ i

have a potential conflict in the time interval [t, t + ∆τ ] prior to any resolution commands.
+ (A |tc ∈ [t, t + ∆τ ]) denote the probability of potential conflict after aircraft
And let Pcon i i,∗

Ai receives a resolution command. Then the relationship

+
Pcon (Ai |tci,∗ ∈ [t, t + ∆τ ]) < Pcon
o
(Ai |tci,∗ ∈ [t, t + ∆τ ]) (23)

should be satisfied for some value of ∆τ ∈ R+ . It is desired that all possibility for a

110
potential conflict disappears after implementing a resolution command, such that

+
Pcon (Ai |tci,∗ ∈ [t, t + ∆τ ]) = 0. (24)

According to the abstraction of conflict-resolution systems provided in Chapter 5, for the

+ (A |tc ∈ [t, t + T + H ]) = 0.
guaranteed conflict-free time HR , Equation 24 becomes Pcon i i,∗ R R

Now, what occurs beyond TR +HR minutes is of particular interest in this chapter. Conflict-

resolution solutions are modeled not to satisfy Equation 24 for ∆τ > (TR + HR ), so there

always exists some probability of secondary conflicts. As it stands, the dynamic graph model

and the abstractions of conflict-detection and resolution tools do not directly indicate how

secondary conflicts are taken into account in a simulation environment. In practice, to

generate a secondary conflict, a specified resolution trajectory is required. However, the

act of generating specified resolution commands is exactly what the abstraction of conflict-

detection and resolution tools aims to avoid. This chapter presents a method for mimicking

secondary conflicts.

6.1 Creation of Complete Conflict-Graphs

To generate traffic scenarios for simulating the abstracted advisory conflict-detection and

resolution tools, the conflict relationships described by the complete uncontrolled graphs

G = (V, E) introduced in Chapter 3 are used. The uncontrolled graph approximates po-

tential conflict relationships that would occur if aircraft flew irrespective of each other and

without any input from an air traffic controller. For the purposes of simulation, the uncon-

trolled graph is expanded to include potential conflicts beyond the standard 5 NM minimum

separation requirement. Instead, all aircraft pairs, (Ai , Aj ), that come within 15 NM and

1000 vertical feet of each other are included as part of the complete edge set E. Also,
miss , as well as the predicted
encoded as part of the graph is the minimum miss-distance Di,j

time of potential conflict, tci,j , given by Equation 3 and Equation 4.

6.2 Simulation Implementation

As noted previously, the use of uncontrolled conflict graphs, G, is not directly applicable to

the dynamic graphs, G(k), modeled in Chapter 4. For simulations, a proxy for the dynamic

111
graph is required because the realization of secondary conflicts is not directly possible with-

out simulating an actual conflict-resolution algorithm. To approximate dynamic graphs, the

uncontrolled conflict graph is incrementally and deterministically sampled. The determin-

istic sampling of the uncontrolled conflict graph adds and removes nodes and edges to the

proxy dynamic graph, thereby mimicking entering and exiting aircraft and the detection

and resolution of potential conflicts.

The simulation procedure to approximate dynamic graphs using uncontrolled graphs is

provided in Procedure 5. In step 1, each traffic scenario is initiated by the uncontrolled

graph G = (V, E). The conflict-detection and resolution process is applied at discrete time-

steps until the simulated time is completed.

At time-step k, the conflict-detection process is applied, as indicated in step 5. Next,

min and D r , the
according to the current time, look ahead value TR , and the parameters Dsep sep
min , D min ), determine
conflict-detection procedure, indicated by the function CD(G, t, TR , Dsep r

all potential conflicts requiring resolution during the time-window [t + TR , t + TR + HD ].

The output of the function is the graph G̃(k). The graph G̃(k) represents an approximation

to the dynamic graph G(k), and is constructed by

Ṽ(k) ={vi |vi ∈ V ∩ Ai ∈ visible and controllable} (25)

miss min
Ẽ(k) ={(ni , nj )|Di,j < (Dsep + (tci,j − t − Tr )Dsep
r
) ∩ (ni , nj ) ∈ V ∩ (ni , nj ) ∈ visible}.

(26)
Aircraft Ai , as represented in the approximated dynamic graph, is considered visible and

controllable if its arrival time, tai , is greater than t + TR . Furthermore, all potential conflicts

occurring after t + TR are visible.

The next steps correspond to the conflict-resolution procedure. Based on the approxi-
˜
mated dynamic graph G̃(k), an approximated decision-graph, Ĝ(k), is constructed to include

only those potential conflicts within HD + TR of the current time. The conflict-decision pol-
˜
icy (e.g. FCFS, MCRT, random) is applied through the function CR(Ĝ(k)) in step 7.

Implicit in the function call is that any other additional information required in deciding

which aircraft to maneuver (e.g. arrival times, flight phase, etc) is included. After applying

the decision policy, the output of the function is the set of maneuvering aircraft M. Based

112
on which aircraft are maneuvered, the uncontrolled conflict-graph, G = (V, E), is updated.

For an aircraft Ai that is issued a resolution command at time t, all future potential conflicts

between [t + TR , t + TR + HR ], represented in the edge set E, are removed. The procedure is

then repeated until the time-span for the traffic scenario, as defined by tsim , is completed.

Procedure 5 Simulating Conflict-Detection and Resolution Process

1: G = (V, E)

2: k = 0, t = 0

3: while t < tsim do

4: k = k + 1, t = t + δt

5: min , D min )
G̃(k) = CD(G, t, Dsep r
˜ ˜ ˜
6: Ĝ(k) = (V̂(k), Ê(k)) where
˜
V̂(k) = Ṽ(k)
˜
Ê(k) = {(ni , ni )|(ni , ni ) ∈ Ẽ(k) ∩ tci,j ≤ (t + TR + HR )}
˜
7: M = CR(Ĝ(k))

8: E r = {(Ai , Aj )|Ai ∈ M, tci,j ≤ t + TR + HR }

9: E = E \ Er

10: end while

The final steps 8 and 9 of the simulation procedure are key to approximating secondary

conflicts. When a resolution command is guaranteed to be conflict-free, the probability

of conflict during the interval [t, t + TR + HR ] is zero. In other words, Equation 24 holds

true for ∆τ = TR + HR . Following TR + HR minutes, potential conflicts represented in

the uncontrolled conflict graph G are maintained even after resolution commands are imple-

mented. In this way, the simulation keeps the probability of conflict outside the time interval

[t, t+TR +HR ] constant when a resolution command is issued. Not only does the probability

of potential conflict remain statistically similar for the open interval (t + TR + H + R, ∞),

but so does the set of aircraft for which potential conflicts might involve. In this sense, the

procedure maintains the relationship

o
Pcon (Ai |tci,∗ ∈ (t + TR + HR , ∞)) = Pcon
+
(Ai |tci,∗ ∈ (t + TR + HR , ∞)) (27)

113
when applying a wide range of conflict-detection and resolution policies, without actually

dictating specific resolution commands.

Resolution commands often take aircraft off their intended paths, either due to altitude

changes or lateral movements. As such, the aircraft maneuvers might cause interactions

that would not have occurred if aircraft maintained their desired routes. However, there

is currently no experimental validation supporting either claim that conflict probabilities

TR + HR minutes after a resolution command remain the same or change. Therefore, the

assumption represented by Equation 27 is maintained.

Note, the implementation described in Procedure 5 is invariant to the value of TR , when

assuming that conflict-free resolution commands of length HR always exists. As such, future

simulations do not use the term.

6.3 Parameter Settings and Traffic Scenarios

Using the simulation described in Procedure 5, a wide variety of parameter settings and

traffic cases are considered in the execution of the abstracted conflict-detection and resolu-

tion tools. In line with the intent of supporting a 3X traffic demand by 2025, simulations

consider 50 traffic scenarios for intensities from 1X to 3X current traffic levels. The sce-

narios are selected from the 100 uncontrolled traffic scenarios generated in Chapter 2, and

parsed down to consider only the 10 hour window between 10AM and 8PM. Explicitly

the traffic intensities, I, come from the set, {1, 1.25, 1.5, . . . , 2.75, 3}X for the three sectors,

ZMP12, ZMP16, and ZMP42. A description of the traffic and a primary conflict analysis

of the considered airspaces was provided Chapter 3. The three sectors chosen for simula-

tion are representative of distinct classes of airspaces commonly found in en route centers.

ZMP42 is a high altitude sector with the most traffic flowing east-bound or west-bound;

the majority of traffic in ZMP16 is beginning its initial descent into Minneapolis St-Paul

International Airport; and ZMP12 is a large high-altitude sector with numerous crossing

routes. Figure 61 illustrates each sector’s location and traffic within the airspace.
min × D r × δt × H × H ), for each traffic
The parameter or configuration space, (I × Dsep sep D R

scenario and decision policy covers a wide range, from best-case to a slightly degraded

114
 Figure 61: Traffic density map of ZMP center, highlighting sectors of interest.

version of current-day operations. The range of values for each parameter are as follows:

• Dsep
min ∈ {4, 5, 6, 7, 8, 9, 10} NM

• Dsep
r ∈ {0, 1/6, 1/3, 1/2} NM/min

• δt ∈ {0, 1, 2, 3, 4} minutes

• HD ∈ {0, 1, 2, 4, 6, 8, 10, 15, 20, ∞} minutes

• HR ∈ {0, 1, 2, 4, 6, 8, 10, 15, 20, ∞} minutes

In addition to the three decision policies described in Chapter 5, the 5 policies in

Appendix A are included as part of the simulations. A total of 45,662,400 simulations were

run, covering a range of airspaces, traffic intensities, decision policies, and configurations.

115
 CHAPTER VII

SIMULATION ANALYSIS

Using the abstracted models in Chapter 5, and the simulation environment in Chapter 6,

the conflict-resolution taskload resulting from advisory conflict-detection and resolution

tools are analyzed. The examination of conflict-resolution taskload focuses on a sensitivity

analysis of the parameters describing the advisory decision-support tools.

The abstraction of conflict-detection systems considers how potential conflicts are iden-
min and D r . The value D min , the aircraft
tified. The corresponding parameters are Dsep sep sep

spacing distance, defines a safety region around each aircraft that no other aircraft is al-

lowed to penetrate. If two safety regions are projected to overlap in time, then a resolution

command is advised to at least one of the aircraft. Uncertainty in trajectory prediction

r , which describes the temporal
and the conflict-detection process is parameterized by Dsep

growth rate of the safety regions.

The abstraction and parameterization of conflict-resolution tools considers the decision

policy used to select which aircraft are maneuvered; how much information is used in the

decision-making process; how long resolution commands are guaranteed conflict-free; and

how often the conflict-resolution problem is solved. Example decision policies include a

simple first-come, first-served policy, a policy that maneuvers randomly selected aircraft,

and an optimal decision policy that uses the least possible number of resolution commands.
min , D r , δt, H , and H ) and changing the
Adjusting each of the parameters (Dsep sep D R

polices, simulations are executed at different traffic intensities. Following the completion

of the simulations, the conflict-resolution taskload is extracted for each traffic scenario and

configuration, and scaled according to the traffic intensity (see Hypothesis 1 in Chapter 2).

Aggregate statistics are computed.

The next section explores basic relationships between each parameter and the conflict-

resolution taskload. A series of sensitivity analysis studies are then performed. The goal is

116
to extract best practices, enabling the design of advisory conflict-detection and resolution

tools.

The first sensitivity analysis seeks to understand the potential benefits of improving a

single parameter. Understanding the potential impact of a single parameter and its behav-

ior with regards to conflict-resolution taskload, helps motivate the subsequent studies. In

Section 7.3, a statistical comparison of decision-policies is performed to understand their

relationships to conflict-resolution taskload. The next section addresses the relative value of

reduced aircraft spacing as compared to improved conflict-resolution tools. Section 7.5 fol-

lows with a study that considers the benefit of long-term guaranteed conflict-free resolution

commands. Finally, the last analysis examines how uncertainty and solve-times should be

managed. Based on results from each of the studies, suggestions for the design of advisory

conflict-detection and resolution tools are provided in Section 7.7.

7.1 Aggregate Behaviors

In this section a preliminary examination of the conflict-resolution taskload is provided to

min , D r , δt, H ,
motivate future studies. Isolating an individual parameter (i.e. I, Dsep sep D

and HR ), the expected conflict-resolution taskload E[N ] is plotted as a function of the

considered term.

Traffic Intensity and Aircraft Spacing. Figure 62 illustrates the quadratic and linear

behavior of the expected conflict-resolution taskload as a function of the traffic intensity,

min . Regardless of the policy, sector, or remaining configu-
I, and the aircraft spacing, Dsep

ration description, the quadratic and linear models hold. Averaging 50 simulation runs for

each configuration and applying a least-squares fit, errors between predicted averages and

simulation averages closely match. Table 4 contains the worst-case average and maximum

errors over all implementations, policies, sectors, and traffic intensities. As indicated in

the table, the quadratic model (E[N ] = cI 2 ) for predicting the conflict-resolution taskload

as a function of traffic intensity has a maximum error of 5.6 resolution commands. Like-
min (E[N ] = cD min ) is always within 1.8 resolution commands
wise, the linear model for Dsep sep

117
 Conflict-Resolution Taskload
Conflict-Resolution Taskload
600
600
500
400
400

200 300

200
0 4 6 8 10
1 1.5 2 2.5 3
min
Dsep [NM]
I
(a) Traffic Intensity (b) Aircraft Separation

Figure 62: Qualitative behavior of conflict-resolution taskload with regards to the traffic
intensity and aircraft spacing.

Table 4: Worst-case model errors between simulation averages and predicted averages over
all configurations, traffic intensities, and sectors.

Parameter Max Abs. Error Mean Abs. Error Max % Error Mean % Error
I 5.6 3.19 8.3% 4.4%
min
Dsep 1.8 0.2 1.1% 0.2%
r
Dsep 11.6 1.3 6.8% 1.4%
δt 4.7 0.4 1.5% 0.3%

of the expected conflict-resolution taskload generated by the simulations. The model de-

scribing the conflict-resolution taskload according to the intensity, I, and aircraft spacing,
min , remains the same with controlled or uncontrolled traffic. Recall that a prior study
Dmsep

in Chapter 3 demonstrated that the number of uncontrolled conflicts can be described by

the model cI 2 Dsep

min , where c is a sector specific constant.

r
Uncertainty and Solve-Times. Isolating the terms Dsep and δt individually, the ex-

pected conflict-resolution taskload grows linearly with each parameter, as shown in Figure 63.
r
The corresponding models predicting the conflict-resolution taskload are E[N ] = c1 +c2 Dsep

and E[N ] = c3 + c4 δt. For some implementations, the slope of the line of best-fit for the

conflict-resolution taskload approaches or equals zero. That is, c2 ∼ 0 or c4 ∼ 0. Later

r
studies indicate that the terms Dsep and δt are coupled. When δt is small, the effect of

uncertainty on the conflict-resolution taskload is small, and vice versa. The linear models

118
 Conflict-Resolution Taskload

Conflict-Resolution Taskload
400 520

500
350

480
300
460
0 0.2 0.4 0 1 2 3 4
r
Dsep [NM/min] δt [min]
(a) Prediction Uncertainty Growth (b) Solve-Time
Rate

Figure 63: Qualitative behavior of conflict-resolution taskload with regards to the uncer-
tainty and solve-time.

match closely to the expected conflict-resolution taskload of the simulations, as indicated

by the small errors in Table 4.

Decision-Horizon Time. The decision-horizon HD , exhibits different behaviors in com-

parison to the other parameters. For the decision-horizon time, when using a rule-based

heuristic policy with conflict-priority ordering, the conflict-resolution taskload is indepen-

dent of HD , as shown by the constant trend-lines in Figure 64a and Figure 64b. The inde-

pendence between HD and the conflict-resolution taskload is apparent through examination

of the procedures describing rule-based policies. For the rule-based heuristic policies with

conflict-priority ordering, each potential conflict is considered independently of others; the

selection of which aircraft to maneuver does not depend on the decision-graph. Thus, the

size of the decision-graph, dictated by HD , does not change which aircraft are maneuvered.

In contrast, the minimum conflict-resolution taskload policy (MCRT) does not exhibit this

property, as shown in Figure 64c. With larger values of HD , the decision-graph increases

in size, thereby providing the policy with more information to make optimal decisions.

There is however a limitation to the benefit of increasing HD . In fact, the reduction in

conflict-resolution taskload, E[N ], decreases as HD → ∞. And furthermore,

δE[N ]
lim →0 (28)
HD →∞ δHD

119
 Conflict-Resolution Taskload

Conflict-Resolution Taskload

Conflict-Resolution Taskload
500 500 500
480 480 480
460 460 460
440 440 440
420 420 420
400 400 400
0 20 40 60 80 ∞ 0 20 40 60 80 ∞ 0 20 40 60 80 ∞
HD [min] HD [min] HD [min]
(a) Random (b) First-Come First-Serve (c) Minimum Conflict-Resolution
Taskload

Figure 64: Qualitative behavior of conflict-resolution taskload with regards to the decision-
horizon time.

Conflict-Free Resolution Time. The plots in Figure 65 demonstrate the value of im-

proving conflict-resolution algorithms through the parameter HR . A larger value of HR

increases the conflict-free resolution time, which prevents secondary conflicts. The plots

represent the random policy and MCRT policy for the same setting and parameter con-
min × D r × δt, H ). For both decision policies, as H increases, the
figuration (I × Dsep sep D R

expected conflict-resolution taskload decreases. Similar to the previous case with HD , there

are diminishing returns on improving the conflict-free time. It is worth noting that for

the same configuration, the conflict-resolution taskload for both the random policy and the

MCRT policy closely match. Minor differences in the conflict-resolution taskload between

the optimal and random polices suggests that the choice in decision-policy has limited effect

on the conflict-resolution taskload. This topic is later explored in Section 7.3.

Complete Conflict-Resolution Taskload Model. Based on the relatively simple re-

lationships indicated by Figure 62 and Figure 63, there is a temptation to formulate the
min , D r , δt) around constant values of H
polynomial model, E[N ] = f (I, Dsep sep D and HR ,

120
 Conflict-Resolution Taskload

Conflict-Resolution Taskload
700 700

600 600

500 500

400 400
0 20 40 60 80 ∞ 0 20 40 60 80 ∞
HR [min] HR [min]
(a) Random (b) Minimum Conflict-Resolution
Taskload

Figure 65: Qualitative behavior of conflict-resolution taskload with regards to conflict-free

resolution time.

where

min r
f (I, Dsep , Dsep , δt) = c1 (I 2 )(Dsep
min r
)(Dsep ) + c2 (I 2 )(Dsep
min r
)(Dmin )(δt)

+ c3 (I 2 )(Dmin
r
) + c4 (I 2 )(Dmin
r
)(δt)
min r min r
+ c5 (I)(Dsep )(Dmin ) + c6 (I)(Dsep )(Dmin )(δt)
min min
+ c7 (I)(Dsep ) + c8 (I)(Dsep )(δt). (29)

A complete model enables prediction of the conflict-resolution taskload for any given traf-

fic scenario and conflict-detection and resolution system. However, besides providing the

conflict-resolution taskload, such a complete model provides little insight. Because traffic in

each sector is unique (as demonstrated in Chapter 3), any best-fit coefficients are not con-

sistent across airspaces. Therefore, study of the best-fit coefficients for the complete model

does not allow for more general conclusions on the relationship between the parameters and

the conflict-resolution taskload. Furthermore, if the conflict-resolution taskload model in

Equation 29 is expanded to consider HD and HR , then a simple polynomial model fails. A

polynomial model is unable to capture the asymptotic behavior of the conflict-resolution

taskload with regards to the parameters HD and HR . Also, the relationships indicated in

Figure 64c and Figure 65 do not correspond to simple functions (e.g. 1/x, e−x ). Certainly,

it is possible to truncate the results over the range HD ∈ [0, 20] minutes and HR ∈ [0, 20]

121
minutes for model fitting. Doing so allow a 4th -order polynomial in HD and HR to describe

the conflict-resolution taskload; lower order polynomials result in a poor fit. A complete

model that considers all parameters, including HD and HR over the truncated domain, re-

sults in a 13 degree polynomial. Such a high-degree polynomial makes drawing conclusions

from the coefficients increasingly difficult, even when focusing on a single sector.

Instead of focusing on mathematical models to prediction the conflict-resolution taskload,

alternative methods for understanding the relative importance of each parameter are pre-

ferred.

7.2 Initial Parameter Sensitivity Analysis

An initial parameter sensitivity analysis was performed on the conflict-resolution taskload

simulation data. The goal of the sensitivity analysis is to understand how improving one

particular parameter can aid in reducing the conflict-resolution taskload. More specifically,

what are the upper and lower bounds of improvement. Along these lines, the following

example question can be asked:

If the conflict-free resolution time is improved from HR =

0 minutes to HR = ∞ minutes, then what is the greatest

expected percent reduction and the smallest expected percent

reduction in the conflict-resolution taskload, regardless of

min × D r × δt × H )?
the remaining configuration (I × Dsep sep D

Replacing HR with another parameter and updating the remaining configuration, answering

a more generic form of the question above provides upper and lower bounds of improvement

according to a single parameter.

To support the sensitivity analysis, some additional notation is required. Let CP denote

the subset of configurations with equal parameter values, excluding the parameter P . So

for all c1 , c2 ∈ CP , where

min r
c1 = (I1 , Dsep,1 , Dsep,1 , δt1 , HD,1 , HR,1 ),
(30)
min r
c2 = (I2 , Dsep,2 , Dsep,2 , δt2 , HD,2 , HR,2 ),

122
it is required that all parameters are equal except the parameter corresponding to P . For

example, CHD corresponds to the space of configurations in which the parameter HD varies,

while all others remain constant.

Following this notation, E[N ]CHD , denotes the expected conflict-resolution taskload as
min , D r , δt, H ,
a function of HD , for a specific configuration with constant values of I, Dsep sep R

and a set decision policy.

For the more generic question, the greatest possible improvement and the smallest pos-

sible improvement for an arbitrary parameter P are given by

 
max(E[N ]cP ) − min(E[N ]cP )
RPmax = max (31)
cP ∈CP max(E[N ]cP )
and
 
max(E[N ]cP ) − min(E[N ]cP )
RPmin = min , (32)
cP ∈CP max(E[N ]cP )
respectively. The values RPmax and RPmin are henceforth referred to as the dynamic range of

improvement for the parameter P .

min , D r , δt, H , and H
The dynamic range for the parameters Dsep sep R D are shown in

Figure 66 and Figure 67 for the FCFS and MCRT policies over the sectors ZMP12, ZMP16,

and ZMP42. All parameters consider the complete range of simulation values (see Section 6.3),
min . Instead, for D min , the potential improvement from changing aircraft spac-
except for Dsep sep

ing from 9 NM to 4 NM is tested to better reflect the change from current operations to

future operations. The dark bars indicate the range of improvement for a single variable. For

example, for the FCFS policy in ZMP12, improvement of the conflict-free time, HR yields a
min = 0% and Rmax = 20%.
maximum reduction in conflict-resolution taskload between RH R HR

The dynamic ranges in Figure 66 and Figure 67 indicate that the largest reduction in

conflict-resolution taskload comes from reducing aircraft spacing. The best-case improve-

ments show that reducing aircraft spacing can decrease the conflict-resolution taskload ap-

proximately 60%, regardless of the sector or decision policy. Moreover, even the lower bound
min is large, (Rmin ∼ 45%), and exceeds the best-case improvement
of improvement for Dsep Dmin sep

for any other parameter.

123
 % Reduction
60 60 60
40 40 40
20 20 20
0 0 0
min δt D r HR HD
Dsep min δt D r HR HD
Dsep min δt D r HR HD
Dsep
sep sep sep

(a) ZMP12 (b) ZMP16 (c) ZMP42

Figure 66: FCFS parameter sensitivity analysis: Range of percent reduction in conflict-
resolution taskload.
% Reduction

60 60 60
40 40 40
20 20 20
0 0 0
min δt D r HR HD
Dsep min δt D r HR HD
Dsep min δt D r HR HD
Dsep
sep sep sep

(a) ZMP12 (b) ZMP16 (c) ZMP42

Figure 67: MCRT parameter sensitivity analysis: Range of percent reduction in conflict-
resolution taskload.

r , and H , show similar

Compared to each other, the remaining parameters δt, Dsep R

reductions in conflict-resolution taskload. However, there are variations between the sectors.

For example, the improvement of increasing HR shows greatest potential benefit in sector
r . That is, Rmax ≥ Rmax ≥ Rmax . However, the ordering of the
ZMP42, over δt, and Dsep HR δt r
Dsep

best-case improvements is not the same for ZMP12 and ZMP16. For the MCRT and FCFS
max ≥ Rmax ≥ Rmax for ZMP12 and ZMP16.
policies, Figure 66 and Figure 67 show that Rδt r
Dsep HR

For both policies, and all sectors, the sensitivity analysis indicates that the decision-

horizon time, HD , has the least impact on the conflict-resolution taskload. In the case of
max = 0, indicating that the decision-horizon has no effect on the conflict-
the FCFS policy, RH D

resolution taskload. This result is the same result illustrated in Figure 64b of Section 7.1.

Perhaps more interesting is that there exist configurations where improvement to a sin-
r , H , and
gle parameter does not reduce the conflict-resolution taskload. For P = δt, Dsep R

HD , the lower bound improvement is RPmin = 0 for both policies across all three sectors.

124
Such a result is surprising. However, it might indicate that at lower traffic levels, the bene-

fits of conflict-resolution tools are limited. Furthermore, optimizing the conflict-resolution

taskload over a subset of parameters allows others to be less than ideal. For example, if

HR = 20 minutes and δt = 0, then an uncertainty growth in trajectory prediction given

r = 0.5 NM/min might be inconsequential to the conflict-resolution taskload. Thus,
by Dsep

such a result might imply that instead of improving trajectory-prediction tools, systems can

be designed to quickly detect and resolve potential conflicts to overcome limitations as a

result of uncertainty.

7.3 Policy Comparison

Results in Section 7.1 indicate the potential for only limited gains when implementing the

MCRT policy over other rule-based heuristic policies (recall the comparison of the conflict-

resolution taskload curves in Figure 65). In this section, a more detailed consideration

of the expected conflict-resolution taskload as a function of the decision policy is taken.

Using statistical tests, A-B comparisons are made to determine the value of minimizing the

conflict-resolution taskload. Ultimately, hypothesis testing demonstrates that the decision

policy has at most a minor effect on the conflict-resolution taskload.

The first set of tests considers the conflict-resolution taskload over the 10 hour time

period for each of the 50 traffic scenario. Using analysis of variance (ANOVA) and Tukey’s

Honestly Significant Difference (HSD) test, multiple comparisons are made to test if there

is a statistical difference between any two decision policies for a set configuration [82]. The

degrees of freedom for Tukey’s HSD test is 392 (8 × 50 − 8). For a specific configuration
min × D r × δt × H × H ), let N A and N B denote the conflict-resolution taskload
(I × Dsep sep D R i i

for traffic scenario i, and decision policies A and B. The mean conflict-resolution taskload is

given by E[N A ] and E[N B ]. The Tukey’s HSD test (α = .05) is applied with the following

hypothesis between each pair of decision policies :

Ho : E[N A ] = E[N B ]

Ha : E[N A ] 6= E[N B ]

When the null hypothesis, Ho , is accepted, then the conflict-resolution taskloads of the

125
 t
(a) Policy A

t
(b) Policy B

t
(c) Policy C

Figure 68: Representative timeline of conflict-resolution advisories.

two policies are similar. Otherwise, acceptance of the alternative hypothesis implies that

the policies result in different conflict-resolution taskloads.

ANOVA and Tukey’s HSD testing requires that the observed conflict-resolution taskload

data come from normal distributions. Additionally, it is required that data samples are

independent from each other, i.e. NiA is independent from NjA for i 6= j. Both properties

are assumed to be satisfied.

As an example, consider the representative timelines in Figure 68 for three arbitrary

decision policies. Each pulse along the timelines indicates that a resolution command is

advised. When the conflict-resolution taskloads are similar, the number of pulses closely

match, indicating no significant difference between the two policies. For the advisory time-

lines of Policy A and Policy B, matched testing would indicate no difference between the

policies’ conflict-resolution taskload. However, if there is a significant difference in the

conflict-resolution taskload, as illustrated by the differences between Policy B and Policy

C, then hypothesis testing rejects the null hypothesis.

The hypothesis testing is applied between all policies, matching traffic configurations

and scenarios. A summary of the test results is provided in Table 5, Table 6, and Table 7

for sectors ZMP12, ZMP16, and ZMP42, respectively. Policies 1-7 correspond to rule-based

heuristic policies, while Policy 8 corresponds to the MCRT policy. In each table entry,

the total number of configurations that are statistically different according to the testing

procedure is listed. As a reference, there are 48,384 configurations for each policy and sector
min , D r , δt, H , and H .
covering a range of values for the settings I, Dsep sep D R

126
 Table 5: Policy Comparisons: ZMP12

Policy 1 2 3 4 5 6 7 8
1 - 0 0 0 0 0 0 1580
2 - - 0 0 0 0 0 2665
3 - - - 0 0 0 0 2665
4 - - - - 0 0 0 2379
5 - - - - - 0 0 1616
6 - - - - - - 0 2382
7 - - - - - - - 2268
8 - - - - - - - -

Table 6: Policy Comparisons: ZMP16

Policy 1 2 3 4 5 6 7 8
1 - 0 0 0 0 0 0 4715
2 - - 0 0 0 0 0 3913
3 - - - 0 0 0 0 3913
4 - - - - 0 0 0 3810
5 - - - - - 0 0 4400
6 - - - - - - 0 4944
7 - - - - - - - 4869
8 - - - - - - - -

A major result of the analysis demonstrates that there is no statistical difference between

any of the rule-based heuristic policies. As shown in the three tables, between the seven

policies, all of the configurations accept the null hypothesis. Based on this result, the

following conjecture is made:

Conjecture 1. Excluding cooperative conflict-resolution, there is no significant difference

in the conflict-resolution taskload between any two rule-based heuristic policies.

Table 7: Policy Comparisons: ZMP42

Policy 1 2 3 4 5 6 7 8
1 - 0 0 0 0 0 0 12322
2 - - 0 0 0 0 0 14261
3 - - - 0 0 0 0 14261
4 - - - - 0 0 0 13861
5 - - - - - 0 0 13298
6 - - - - - - 0 13508
7 - - - - - - - 12937
8 - - - - - - - -

127
 8 8 8

% Reject Ho

% Reject Ho

% Reject Ho
6 6 6
4 4 4
2 2 2
0 0 0
1 1.5 2 2.5 3 1 1.5 2 2.5 3 1 1.5 2 2.5 3
Traffic Intensity Traffic Intensity Traffic Intensity
(a) ZMP12 (b) ZMP16 (c) ZMP42

Figure 69: Percent of configurations for which the MCRT policy is statistical different from
a rule-based heuristic policy (Reject Ho ).

The hypothesis testing reveals that in many cases there is a statistically relevant dif-

ference between the seven rule-based heuristic policies and the MCRT policy (see the last

column in each table). However, this statistical difference does not exist for all configura-

tions. In fact, the number of configurations that accept the alternative hypothesis grows

with traffic intensity. Figure 69 illustrates this point. Each line in the plots represents one

of the seven rule-based heuristic policies. The trend lines show that as the traffic inten-

sity increases, a greater fraction of configurations become statistically different from the

MCRT policy. While not shown, a similar trend exists as the aircraft spacing distance

increases. The implication of both results is that the MCRT policy becomes progressively

more effective as the number of potential conflicts increases.

While the MCRT policy is deemed statistically different from the other policies, the

exact benefit is not yet stated. Figure 70 illustrates the maximum difference in expected

conflict-resolution taskload between the MCRT policy and the random policy. Each line
r , and considers
corresponds to a subset of configurations parametrized by the value Dsep

every possible implementation of (δt × HD × HR ) at the 3X traffic level. According to

the results in Figure 70, for ZMP12 and ZMP16, the largest average difference in conflict-

resolution taskload between the MCRT policy and the random policy is 9 less maneuvers

over a 10 hour time period. For ZMP42, the reduction is just over 10 when aircraft are

spaced at 6NM. Such a reduction only accounts for an average difference of 1 less maneuver

per hour. Considering that there is far greater variance in conflict-resolution taskload

128
 max(N 6 − N 8 )%

max(N 6 − N 8 )%
20 20

10 10

0 0
4 6 8 10 4 6 8 10
min
Dsep min
Dsep
(a) ZMP12 (b) ZMP16
max(N 6 − N 8 )%

r
Dsep = 0NM/min
20
r
Dsep = 2/12NM/min
10 r
Dsep = 4/12NM/min
r
Dsep = 6/12NM/min
0
4 6 8 10
min
Dsep
(c) ZMP42

Figure 70: Maximum difference in the expected conflict-resolution taskload between the
MCRT policy and the random policy over a range of configurations.

per hour (5 ≤ σ 2 ≤ 30), it is unlikely an air traffic controller would notice such a small

reduction in the conflict-resolution taskload. At lower traffic intensities, the differences

in the conflict-resolution taskload, even when statistically different, are marginal. Overall,

when comparing the maximum difference in conflict-resolution taskload between the MCRT

policy and any other policy, the potential benefits appear limited.

A similar analysis can be extended to shorter time-windows of 2.5 minutes. While 10

less resolution commands over 10 hours may appear insignificant, if the deductions are

concentrated during a short-burst of activity, then there might be some benefit of the

MCRT policy in reducing controller workload. Figure 71 contains example distributions of

the number of resolution commands used to space aircraft during the busiest 2.5 minute
min = 6,
time periods for the 1X, 2X, and 3X traffic levels over ZMP42 (configuration: Dsep
r
Dsep = 0, δt = 0). Using a two-sample Kolmogorov-Smirnov goodness-of-fit test (KS

test) with α = .01, the conflict-resolution taskload distributions are tested to see if they

come from different populations. Accordingly, the following null hypothesis and alternative

129
 Probability

Probability

Probability
0.8 0.8 0.8
0.6 FCFS 0.6 0.6
0.4 MCRT 0.4 0.4
0.2 0.2 0.2
0 0 0
012345678 012345678 012345678
# of Commands # of Commands # of Commands
(a) 1X: Accept Ho (b) 2X: Accept Ho (c) 3X: Reject Ho

Figure 71: Probability distributions of the conflict-resolution taskload for the busiest 2.5
minute time period in each traffic scenario.

hypothesis are set:

Ho : Samples are drawn from the same distribution

Ha : Samples are drawn from different distributions

If the null hypothesis is accepted, then there is no distinction in the conflict-resolution

taskload between the two policies.

Application of the KS test reveals that there is no statistical difference between the

MCRT policy and the random policy at the 1X and 2X traffic intensities for the config-

uration corresponding to Figure 71. However, at the 3X traffic level, the null hypothesis

is rejected, implying that there is a difference between the distributions of the conflict-

resolution taskload. Figure 71c illustrates that the distribution of the conflict-resolution

taskload for the MCRT policy is skewed to the left in comparison to the random policy.

When applying the same testing procedure to ZMP12 and ZMP16, there are no differ-
min ≤ 6 NM, regardless of the traffic intensity.
ences found for any configurations with Dsep

In the case of ZMP42, the only configurations for which the MCRT policy outperforms the

random policy is when the solve-times is δt = 0 minutes. A complete list of configurations

min ≤ 5
for which the null hypothesis of the KS test is rejected at the 3X traffic level, when Dsep

NM, is provided in Table 8. Also included as part of the table is the probability that the

MCRT policy reduces the conflict-resolution taskload by at least two resolution commands.

For most cases when the null hypothesis is rejected, the guaranteed conflict-free time,

HR , is greater than or equal to 15 minutes. Likewise, the decision-horizon times are large.

By current standards of performance, HR = 15 minutes is quite large, especially when one

considers that the look ahead time for potential conflicts is typically greater than 5 minutes

130
Table 8: Configurations with Dsepmin ≤ 5N M for which there is a statistical difference in the

conflict-resolution taskload distribution over 2.5 minutes for ZMP42

min [NM] r [NM/min] δt [min] HD [min] HR [min] P (NiA − NiB  ≥ 2)
 
Dsep Dsep
5 0 0 10 15 0.18
5 0 0 15 15 0.10
5 0 0 15 20 0.16
5 0 0 15 100 0.10
5 0 0 20 100 0.14
5 0.167 0 15 20 0.18
5 0.167 0 15 100 0.16
5 0.167 0 20 100 0.12
5 0.333 0 15 15 0.10
5 0.500 0 8 10 0.12
5 0.500 0 10 20 0.12
5 0.500 0 15 20 0.14

(i.e. TR ≥ 5 minutes). State-of-the-art trajectory prediction tools are optimistically capable

of detecting potential conflicts 20 minutes into the future [5]. It is also worth noting that

the solve-time of δt = 0 refers to event-based conflict-detection and resolution implemen-

tation. For event-based implementations, as soon as a potential conflict is identified, it is

immediately resolved. For many conflict-resolution algorithms, the ability to quickly gener-

ate resolution commands for aircraft, especially when the decision-horizon and conflict-free

times are large, poses quite a challenge.

According to Table 8, there is only a marginal fraction of traffic scenarios (< 0.20) for

which the MCRT policy reduces the number of advisory commands by two or more. Addi-

tional analysis indicates that between 35%-53% of the traffic scenarios show no improvement

in conflict-resolution taskload across the configurations.

min ≤ 5 NM
When considering traffic intensities less than 3X, there are no cases with Dsep

and HR ≤ 20 minutes for which the MCRT policy has a statistically different conflict-

resolution taskload distribution from the random policy. Thus it can be argued that the

MCRT policy does not pose a substantial benefit over any other rule-based heuristic policy.

Differences between the other rule-based polices and the MCRT policy reflect the find-

ings above. Furthermore, extending the KS test to the other rule-based policies, there is no

statistical difference found between them.

131
 Working with the assumption that improvements in position sensing and trajectory
min ≤ 5 NM)
prediction, enabled by ADS-B, will allow for reductions in aircraft spacing (Dsep

and uncertainty, and that complete airspace redesigns will occur before 3X traffic, an even

stronger conjecture can be made:

Conjecture 2. There are no significant differences in the conflict-resolution taskload be-

tween any two decision-policies (excluding cooperative conflict-resolution).

The previous conjecture can be applied to understanding the value of the decision-

horizon time, HD . As it stands, the MCRT policy is the only policy that benefits by

extending the decision-horizon time. In contrast, an increase in the value of HD has no

effect on conflict-resolution taskload of the rule-based heuristic policies with conflict-priority

ordering. Differences in the conflict-resolution taskload between the MCRT policy and any

other policy is small for the same configuration (regardless of the value of HD ). Thus, it can

be stated that any increase in HD , while potentially resulting in statistical differences over

10 hour time windows, does not effectively reduce the conflict-resolution taskload. Hence

the corollary below.

Corollary 1. The air traffic controller is unlikely to detect any benefit in reduced conflict-

resolution taskload due to an increase in the decision-horizon time, HD .

The major result of this section is that the decision policy and the decision-horizon

time has limited or imperceivable effects on the conflict-resolution taskload of an advisory

decision-support tool.

min
7.4 Relative value between Dsep and HR

The preliminary results in Section 7.2 indicate that a decrease in aircraft spacing can yield

far greater reductions in conflict-resolution taskload in comparison to adjusting other pa-

rameters. As it stands, next generation air traffic systems (e.g. NextGen and SESAR) all

include the introduction of Automatic Dependent Surveillance-Broadcast (ADS-B) to report

aircraft positions using global positioning systems (GPS). Furthermore, improved trajec-

tory prediction and 4D trajectory-based operations are considered imperatives to managing

132
congestion and increasing throughput, especially in terminal areas [38, 46, 90]. Researchers

hypothesize that implementing ADS-B and improving trajectory predictions will allow for

closer aircraft spacings [31, 81]. Thus, it is likely that the reductions in conflict-resolution

taskload identified in Section 7.2 will one day be realized. However, there are still unan-

swered questions concerning the relative importance of reducing aircraft spacing in compari-

son to the introduction and improvement of conflict-resolution algorithms. More specifically,

how should a research and development portfolio be developed to provide equitable support
min is compared to
according to potential benefits? In this section the relative value of Dsep

HR in reducing conflict-resolution taskload.

The value of improved trajectory prediction and position measurement systems enabling
min . Larger values of D min ,
a reduction in aircraft spacing is tested through the parameter Dsep sep

beyond minimum required separations (e.g. 5 NM in current-day operations), represent

increased uncertainty in the current and future positions of aircraft. This uncertainty often

leads to excessive labeling of potential conflicts and to issuing resolution commands that

promote overly-safe and conservative actions by air traffic controllers. The importance

of conflict-resolution algorithms is evaluated through the parameter HR . Increasing the

value of HR simulates improving the capability of conflict-resolution algorithms to ensure

conflict-free trajectories for longer periods of time.

Because the previous section indicates that the value of HD and the underlying decision

policy is insignificant to determining the conflict-resolution taskload, this study strictly

considers the FCFS policy and the configuration setting HD = 0 and δt = 0. Later
r
in Section 7.6, analysis demonstrates that the value of Dsep does not affect the conflict-

resolution taskload when δt = 0. As such, the uncertainty parameter is arbitrarily set to

r = 0 in this section.
Dsep

Applying the FCFS policy to 10 hours of traffic in ZMP42, the expected conflict-

resolution taskload counts for three traffic intensities (1X, 2X, 3X) are depicted in Figure 72.

As the traffic intensity increases, so does the expected conflict-resolution taskload across

all values of HR and Dsep ; each surface representing a single traffic intensity is strictly

greater than the previous. Figure 72 clearly illustrates that as aircraft spacing is reduced,

133
Figure 72: Expected number of advisory resolution commands for the FCFS policy at 1X,
2X, and 3X traffic intensities.

the conflict-resolution taskload decreases. Also, increasing the conflict-free time, HR , for

the conflict-resolution algorithm can significantly reduce the conflict-resolution taskload,

albeit with diminishing returns.

min and H
To begin the study comparing the relative value of Dsep R to the conflict-

resolution taskload, a reference standard is required. Assume that air traffic controllers

(with or without assistance from conflict-resolution tools) are able to issue resolution com-

mands that are guaranteed conflict-free for an average of 4 minutes beyond the look ahead

time TR . Thus, if a potential conflict is identified to occur in 5 minutes, and it is immediately

resolved, then the associated resolution command ensures that the aircraft is conflict-free

for 9 minutes. Depending on the flight-plans of other aircraft, there is the potential for

resolved trajectories to be conflict-free even longer.

The relative benefits of aircraft spacing compared to the advancement of conflict-resolution

algorithms is assessed by determining the value of HR required to achieve the same decrease

in conflict-resolution taskload when aircraft spacing is reduced. For example, consider the

following question:

134
 Let improved trajectory prediction and reporting allow aircraft spac-
min , from D initial to D f inal . For resolution
ing to decrease by ∆Dsep sep sep

commands that are guaranteed to be conflict-free for HR = 4 minutes,

the expected conflict-resolution taskload decreases from E[N initial ] to

E[N f inal ]. What is the required conflict-free time, HR , to achieve the

same decrease in conflict-resolution taskload when aircraft spacing is

initial = D f inal )?
held constant (i.e. Dsep sep

By evaluating the conflict-resolution taskload of the simulations and interpolating be-

tween results, the above question can be answered for a wide range of traffic intensities and

initial and final aircraft spacings. These results are illustrated in Figure 73 and Figure 74

for sectors ZMP12 and ZMP42. As an example, from Figure 74b, to match the decrease

in conflict-resolution taskload associated with a 0.5 NM reduction in aircraft spacing, from

initial = 8 NM to D f inal = 7.5 NM, at the 2.5X traffic intensity, a guaranteed conflict-free
Dsep sep

time of HR = 9.5 minutes is required.

The figures illustrate that in order to match greater reductions in aircraft spacing,

increasingly larger values of HR are needed. For example, at the 3X traffic intensity, to

match the decrease in conflict-resolution taskload when aircraft spacing is reduced by a

initial = 8 NM), a conflict-free time of H = 6 minutes is needed. However, for
0.25 NM (Dsep R

a larger decrease of 1 NM, a 20 minute conflict-free time is required.

Furthermore, as the reduction in aircraft spacing increases, from 0.25 NM to 1 NM, the

space of intensities and initial aircraft spacings for which a matching value of HR exists

grows smaller. Notice that from Figure 73a to Figure 73d (and Figure 74a to Figure 74d),

the no-solution whitespace grows larger and larger.

When slower solve-times are present for the conflict-detection and resolution process,

the value of HR becomes more relevant. Comparing Figure 75a and Figure 75b, for a slower

solve-time given by δt = 2 minutes, the set of intensities and aircraft spacings for which

improving the conflict-free time can match spacing reductions is larger than the δt = 0 case.

This is also true when uncertainty is introduced, as illustrated in Figure 75d. However, that

is not to say that increasing δt is beneficial. On the contrary, Section 7.1 shows that a greater

135
 10 20 10 20
15 15

DInitial

Initial
8 8

sep

Dsep
6 10 6 10

4 5 4 5
1 2 3 1 2 3
Traffic Intensity Traffic Intensity
(a) Spacing reduction of 0.25 NM. (b) Spacing reduction of 0.50 NM.

10 20 10 20
15 15
DInitial

Initial
8 8
sep

Dsep
6 10 6 10

4 5 4 5
1 2 3 1 2 3
Traffic Intensity Traffic Intensity
(c) Spacing reduction of 0.75 NM. (d) Spacing reduction of 1.0 NM.

Figure 73: Required value of HR to match a reduction in aircraft spacing (ZMP12).

10 20 10 20
15 15
DInitial

Initial

8 8
sep

Dsep

6 10 6 10

4 5 4 5
1 2 3 1 2 3
Traffic Intensity Traffic Intensity
(a) Spacing reduction of 0.25 NM. (b) Spacing reduction of 0.50 NM.

10 20 10 20
15 15
DInitial

Initial

8 8
sep

Dsep

6 10 6 10

4 5 4 5
1 2 3 1 2 3
Traffic Intensity Traffic Intensity
(c) Spacing reduction of 0.75 NM. (d) Spacing reduction of 1.0 NM.

Figure 74: Required value of HR to match a reduction in aircraft spacing (ZMP42).

136
 10 20 10 20
15 15

DInitial

Initial
8 8

sep

Dsep
6 10 6 10

4 5 4 5
1 2 3 1 2 3
Traffic Intensity Traffic Intensity
r r
(a) δt = 0min, Dsep = 0 NM/min (b) δt = 2min, Dsep = 0 NM/min

10 20 10 20
15 15
DInitial

Initial
8 8
sep

Dsep
6 10 6 10

4 5 4 5
1 2 3 1 2 3
Traffic Intensity Traffic Intensity
r r
(c) δt = 0min, Dsep = 4/12(d) δt = 2min, Dsep = 4/12
NM/min NM/min

Figure 75: Required value of HR to be equivalent to a reduction in the separation require-

min = 0.5 NM (ZMP42).
ments of ∆Dsep

value of δt results in increased conflict-resolution taskload. The figures imply that if large

uncertainty is present and solve-times are slow, then both a reduction in aircraft spacing

and improved conflict-free times help to decrease the conflict-resolution taskload.

Overall, the results indicate that in order to reduce conflict-resolution taskload, the

most effective approach is to reduce aircraft spacing. While a reduction in spacing is not a

straightforward achievement, the introduction of ADS-B as part of next generation air traffic

systems will surely aid in limiting sensor measurement noise in the position information

provided to controllers. Additionally, the introduction of 4D trajectory-based operations

in conjunction with performance standards to ensure aircraft stay on desired paths will

provide a framework to improve trajectory projections.

Keeping all other variables constant, reductions in the effective aircraft spacing has a

significant impact on conflict-resolution taskload. In fact, for the ZMP42 simulations, the

reduction is linear over all intensities and conflict-free times HR , as shown in Figure 72.
initial = 9 NM, the percent reduction in
When compared to a reference standard of Dsep

137
Figure 76: Percent decrease in the conflict-resolution taskload by reducing aircraft spacing
initial = 9 NM. (At the 1X, 2X, and 3X traffic intensities)
when compared to Dsep

conflict-resolution taskload can be quite significant. According to Figure 76, the number of

potential conflicts can be reduced approximately 60% by decreasing spacing standards to

4 NM, regardless of the value of HR . This improvement is consistent for all traffic intensities,

airspaces, and configurations.

The research presented here suggests that the consequences of improved trajectory pre-

diction and position sensing for conflict detection can result in greater decreases in controller

effort (implicitly through the reduction in conflict-resolution taskload) than the improve-

ment of advisory systems for conflict-resolution. That is not to say that advisory conflict-

resolution systems are not valuable, but rather, they are not the most promising approach

for reducing conflict-resolution taskload. Thus, research, development, and policy efforts

should place greater focus on gaining universal presence of ABS-B on all aircraft, and

on improving aircraft trajectory predictions. In the next section, the benefits of conflict-

resolution systems are studied to better understand what might be considered a satisfactory

conflict-free resolution time.

138
 Conflict-Resolution
min
Dsep =4
600 600 600

Taskload
min
Dsep =6
400 min
400 400
Dsep =8
200 min
Dsep = 10 200 200
0 0 0
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
HR HR HR
(a) 1X Traffic (b) 2X Traffic (c) 3X Traffic

r = 0).
Figure 77: Conflict-resolution taskload for ZMP42 (FCFS, δt = 0, Dsep

7.5 Benefit of Improved Conflict-Resolution

While not nearly as effective as reducing aircraft spacing, the introduction and improvement

of conflict-resolution algorithms can certainly reduce the conflict-resolution taskload. How-

ever, the potential gains are limited. Therefore, instead of striving to achieve the perfect

conflict-resolution algorithm that can guarantee conflict-free travel for indefinite periods,

it is advisable to design simple and robust algorithms that achieve a satisfiable conflict-

resolution taskload. The study in this section focuses on the diminishing returns associated

with the parameter HR .

The conflict-resolution taskload for the FCFS policy with δt = 0 minutes is illustrated

in Figure 77 for different traffic intensities and aircraft spacings. As the traffic intensity and

aircraft spacing increases, the conflict-resolution taskload grows, regardless of the value of

HR . However, for a fixed traffic intensity and aircraft spacing, as the conflict-free time is

increased, the controller receives fewer advisory resolution commands. Similar behavior is

exhibited in sectors ZMP12 and ZMP16 as well.

Figure 78 illustrates the potential benefits of increasing the conflict-free time according

to a reference standard of HR = 6 minutes. For example, for a spacing requirement of

6 NM, improving the guaranteed conflict-free resolution time to 20 or more minutes yields

a maximum reduction in conflict-resolution taskload of 8.5% at the 3X intensity level. This

decrease corresponds to an average of 27 less resolution commands issued over the 10 hours

time period. Smaller reductions occur at lower traffic intensities.

139
 10 10 10

% Reduction
0 0 0
min
Dsep =4
−10 −10 −10 min
Dsep =6
min
Dsep =8
−20 −20 −20 min
Dsep = 10
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
HR HR HR
(a) 1X (b) 2X (c) 3X

Figure 78: Percent reduction in conflict-resolution taskload compared to HR = 6 minutes

(ZMP42).

min = 4 NM, the largest improvement from an increase in H is about a 6.1%

At Dsep R

decrease in conflict-resolution taskload, corresponding to an average reduction of 13 resolu-

tion commands over 10 hours. The decrease corresponds to 1 less resolution command every

50 minutes. From the perspective of the air traffic controller, this difference is probably

imperceptible, especially because the variance in the conflict-resolution taskload within one

hour is far greater. Similarly, diminishing returns exist for ZMP12 and ZMP16. Figure 79
min = 6 NM, improvement to the conflict-
illustrates that even at the 3X traffic level, for Dsep

resolution algorithm only reduces the conflict-resolution taskload by 10 commands (≤ 8%).

This analysis demonstrates that the benefit of increasing the conflict-free time to reduce

controller taskload has diminished returns, especially for HR > 20 minutes. One reason

for the limited improvement associated with significant increases in HR is that sector sizes

are finite. Sectors are also designed to have intersecting regions spaced far apart from each

other.

The distributions of aircraft transit times through the sectors are shown in Figure 80.

For ZMP42, 65% of all aircraft take 20 minutes or less to pass through the airspace, and

virtually all take less than 25 minutes. For ZMP16, the vast majority of aircraft exit the

airspace within 10 minutes. In the context of transit times, the diminishing returns for

HR ≥ 20 minutes for ZMP42 makes sense. Likewise, for ZMP16, because transits times

tend to be less than 10 minutes, Figure 79b and Figure 79d illustrates that there is little

140
 20 20
0 0

Reduction

Reduction
min
Dsep =4
−20 min
Dsep =6 −20
min
Dsep =8
−40 −40
min
Dsep = 10
−60 −60
0 10 20 30 40 0 10 20 30 40
HR HR
(a) 3X: ZMP12 (b) 3X:ZMP16

20 20
% Reduction

% Reduction
0 0

−20 −20
0 10 20 30 40 0 10 20 30 40
HR HR
(c) 3X: ZMP12 (d) 3X:ZMP16

Figure 79: Relative reductions in conflict-resolution taskload for ZMP12 and ZMP16.

benefit to increasing HR beyond this time.

The sector ZMP12 is a bit different from ZMP16 and ZMP42. Because of its size, a

large portion of traffic takes more than 20 minutes to traverse the airspace, as shown in

Figure 80a. However, the plots in Figure 79 show that the benefit of extending HR beyond

15 or 20 minutes is relatively small.

Understanding why a conflict-free time of HR = 15 minutes is satisfactory for ZMP12 can

best be understood through analysis of the inter-arrival conflict times. Figure 81 contains

the distribution of inter-arrival times between an aircraft’s first potential conflict and its

last potential conflict within a sector. In the case of ZMP12, 91% of inter-arrival times are

less than 20 minutes, despite the fact that only 50% of all aircraft traverse the sector in the

same time period.

Comparing the three sectors, 90% of all inter-arrival conflict times occur within 20

minutes, 10 minutes, and 16 minutes for ZMP12, ZMP16, and ZMP42, respectively. Because

each sector is designed according to a specific purpose, there is no one-size fits all value of

141
 0.15 0.15 0.15

Probability
0.1 0.1 0.1
0.05 0.05 0.05
0 0 0
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
Transit Time Transit Time Transit Time
(a) ZMP12 (b) ZMP16 (c) ZMP42

Figure 80: Distribution of aircraft transit times [minutes] through sector (1 minute incre-
ments).

0.1 0.1 0.1

Probability

0.08 0.08 0.08

0.06 0.06 0.06
0.04 0.04 0.04
0.02 0.02 0.02
0 0 0
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
Time between conflicts Time between conflicts Time between conflicts
(a) ZMP12 (b) ZMP16 (c) ZMP42

Figure 81: Distribution of inter-arrival times [minutes] between the first and last potential
conflict of an aircraft (1 minute increments).

HR , however even with a value of HR = 6 minutes the majority of secondary conflicts can

be prevented.

7.6 Managing Uncertainty and Solve-Times

Uncertainty in the conflict-detection process can pose quite a challenge to air traffic con-

trollers and advisory conflict-resolution systems. As a precautionary measure, the need

to manage future workload sometimes prompts air traffic controllers to resolve potential

conflicts that are unlikely to ever become realized [52]. In the case of future systems, with

the introduction of trajectory-based operations and automated hand-offs, many routine and

mundane tasks are offloaded to computer systems. This transference of function provides

the air traffic controller with more time to dedicate to resolving potential conflicts. However,

there is still a need to answer how often conflict-resolution problems should be resolved when

uncertainty is present, and furthermore, when is it beneficial to sacrifice rapid solve-times

for improved conflict-free resolution times.

142
 Conflict−Resolution

Conflict−Resolution
330 330

Taskload

Taskload
300 300
270 270
4 3 0.5 4 3 0.5
2 1 0.25 2 1 0.25
0 0 0 0
δt Drsep δt Drsep
(a) FCFS (b) MCRT

Figure 82: Conflict-resolution taskload as a function of the solve-time and uncertainty

parameters.

For systems that simultaneously solve the conflict-resolution problem for multiple air-

craft, computation times can grow quite large [100]. These slow solve-times prevent rapid

information updates and imply that decisions are made on old information. In the case of

uncertainty, when computation time prevents event-based resolution of potential conflicts,

advisory systems are required to plan in advance and hedge accordingly - much like human

air traffic controllers. Conservative behavior then results in an increase in the number of

advisory resolution commands.

r ,
Study of the solve-times and uncertainty, parameterized by the terms δt and Dsep
r
demonstrate a coupling as shown in Figure 82. When either δt or Dsep is small, the effect

of the other term on the conflict-resolution taskload is diminished. For a given airspace
min , H , H ),
and decision-policy, with fixed values of traffic intensity and configuration (Dsep D R

the expected conflict-resolution taskload over 10 hours can be modeled as a function of the
r
uncertainty and solve-time parameters Dsep and δt:

r r
E[N ] = c1 Dsep δt + c2 δt + c3 Dsep + c4. (33)

r δt.
Note the coupling term Dsep

Using least-squares optimization to solve for the coefficients, the error of the model is

bounded by 2% for all configurations, regardless of policy, airspace, intensity, or remaining

parameters.

For all considered policies, except MCRT, c3 = 0. As long as the conflict-detection and

143
resolution problem is solved in an approximate event-based implementation, there is little

effect of uncertainty on the conflict-resolution taskload. For the MCRT policy, the coefficient

c3 is non-zero but bounded between 0 < c3 < 12. Even with an event-based implementation

of the MCRT policy, uncertainty has a minor effect on the conflict-resolution taskload.

The additional conflict-resolution taskload is a consequence of the MCRT policy making

decisions on uncertain potential conflicts that do not become realized. In the worst case
r
scenario, for δt = 0 and Dsep = 0.5 NM/min, the maximum average increases in expected

conflict-resolution taskload is 6 more resolution commands during the 10 hour simulation

period.

In current operations, there already exist conflict-resolution algorithms capable of solv-

ing even complex conflict configurations in under a minute. [28] and [32] are two such

examples taking different approaches. In [28], the authors make use of genetic algorithms

to simultaneously generate resolution commands for sets of aircraft (# of aircraft ≤ 15).

Aircraft trajectories are represented by a genes sequence representative of motion primi-

tives (i.e. heading, speed or altitude changes). The conflict-resolution algorithm in [28]

can also be implemented sequentially, resolving potential conflicts individually in order of

occurrence [4]. Sequential conflict-resolution degrades the problem to path planning for

a single aircraft in a dynamic environment. Relative to simultaneous conflict-resolution,

sequential path planning can be solved faster for the complete set of aircraft. While not

always capable of generating feasible solutions for both simultaneous and sequential im-

plementations, there is no reason why the gene library of potential trajectory solutions

cannot be expanded to overcome this limitation. As it stands, the gene library does not

consider complicated closed-loop trajectories with multiple lateral off-set, speed, and al-

titude changes. Furthermore, with the advent of multi-core and graphics processing unit

computing, genetic algorithms can take advantage of parallelized implementations to ensure

that feasible solutions can be found, even when large libraries are used [43]. Likewise, the

rule-based heuristic method proposed in [32] can take a similar approach.

For present-day traffic levels, there exist conflict-resolution formulations that can be

solved within 1 minute or in an event-based implementation. However, as traffic intensities

144
and conflict complexities increase, the issue of solve-times will become more pertinent. It is

certainly true that faster solves-times are preferred, as longer solve-times have the potential

to create ancillary problems. For example, in the case of a system-crash or the rejection of an

advisory command, time is needed either to recover the system or to request a new advisory.

A more problematic circumstance arises when the air traffic controller is unaware of a system

failure and continues to wait for advisory resolution commands. For these cases, systems

must be designed to reboot quickly, and generate any necessary resolution commands before

problems occur. Therefore, it is assumed and asserted that conflict-resolution tools must

be capable of resolving the conflict-resolution problem in under 1 minute to ensure they are

useful to air traffic controllers.

The primary question then becomes: Is it preferred to have rapid solution-times (i.e.

δt = 0 minutes) to help manage uncertainty, or is it preferred to sacrifice solve-times to

ensure longer conflict-free times, HR ?

Similar to the benefit analysis provided in Section 7.4, solve-times and conflict-free times

are now compared. In this case, the following question is asked:

Let improved computation (through hardware or software implemen-

tation) result in faster solve times from δt = 1 minute to δt ∼ 0

minutes. Additionally, let the same conflict-resolution system be able

to guarantee conflict-free trajectories for HR = 6 minutes. The im-

provement to a faster solve-time limits the effect of uncertainty, re-

sulting in a decrease in the expected conflict-resolution taskload from

E[N initial ] to E[N f inal ]. If instead solve-times are held constant,

but efforts are taken to strengthen the conflict-resolution algorithms,

what is the required conflict-free time, HR , to achieve the same de-

crease in conflict-resolution taskload?

Figure 83 answers the question for sectors ZMP12, ZMP16, and ZMP42 at different
r
traffic intensities according to the uncertainty parameter Dmin (FCFS decision policy).

145
 20 20 20
1.0X
2.0X

HR [min]

HR [min]

HR [min]
15 15 3.0X 15

10 10 10

0 0.2 0.4 0 0.2 0.4 0 0.2 0.4

r
Dsep [NM/min] r
Dsep[NM/min] r
Dsep [NM/min]
(a) ZMP12 (b) ZMP16 (see text) (c) ZMP42

Figure 83: Required value of HR to match faster solve-times.

r = 0.25 NM/min, a conflict-

Using Figure 83c as an example, at the 2X traffic level, if Dsep

free time of at least HR ∼ 12 minutes is required to equal the reduction in the conflict-

resolution taskload due to an improvement in the solve-time. In some cases, for specific

values of traffic intensity or uncertainty, no value of HR can match the change in conflict-

resolution taskload. This is reflected by the lack of a solution (i.e. line) in Figure 83b; there

is no matching HR for the 1X traffic intensity, and at the 2X and 3X traffic levels, only a

limited range of uncertainty values have solutions. Consistent across the three sectors, as

traffic intensity increases, smaller improvements to HR are required.

Problematic with this analysis is that there is no consistent answer to which is pre-

ferred: faster solve-times or improved conflict-free times. As Figure 83 demonstrates, the

equivalent value of HR is different for each sector. Furthermore, changes in the required HR

between traffic intensities are not consistent. For example, for ZMP42, there is a significant

separation in the equivalent values of HR between the 2X and 3X traffic intentsity lines.
R ≤ 0.33 [NM/min], there is little distinction between the
However, for ZMP12 when Dsep

2X and 3X traffic intensity lines.

Ultimately, there does not exist a standard best practice for the selection of HR and δt.

Rather, each parameter setting needs to be tuned according to the traffic within the sector

and the uncertainty in trajectory prediction. However, considering that uncertainty in

identifying potential conflicts is a function of the trajectory prediction tools, an alternative

r provided large decreases in
approach is to improve these systems. Just as reductions in Dsep

146
 400

Conflict-Resolution
Taskload
min = 4 NM
Dsep
min = 6 NM
Dsep
200 min = 8 NM
Dsep

0
1 2 3 4 5
Step

Figure 84: Representative summary of analysis.

the conflict-resolution taskload when compared to HR , again, the conflict-detection side of

the advisory tools can be improved to make up for short comings in the conflict-resolution

algorithms.

7.7 Review

The results in this chapter can best be understand through Figure 84, which follows a

sequence of steps corresponding to reductions in the performance capabilities of an advisory

conflict-detection and resolution system. Step 1 represents the best-case implementation

min : the decision policy is MCRT, D r
of an advisory system for a given value of Dsep sep =

0 NM/min, δt = 0 minutes, HD = ∞ minutes, and HR = ∞ minutes. In this case, on

average the least number of advisory resolution commands is generated. Step 2 reflects a

change in the decision policy from MRCT to FCFS. With the change, there is an increase

in the expected conflict-resolution taskload. In Step 3, the conflict-free time is reduced

to 6 minutes, which brings with it another increase. In Step 4 and Step 5, the solve-
r
time is slowed to δt = 1 minute, and the uncertainty parameter is changed to Dsep =

0.167 NM/min. The last configuration is representative of an advisory system that could

realistically be implemented.

The most noticeable characteristic of the figure is that the conflict-resolution taskload
min . When D min = 4 NM, even the
is most sensitive to the aircraft spacing parameter Dsep sep

degraded performance setting of the advisory system results in a lower conflict-resolution

min = 6 NM.
taskload than the best-case implementation for an aircraft spacing of Dsep

147
 Therefore, to reduce conflict-resolution taskload, the most effective measure is to de-

crease the aircraft spacing requirement. As stated before, the introduction of ABS-B will

most likely go a long way in enabling this reduction. Furthermore, improved trajectory

prediction allows for greater certainty in labeling potential conflicts. In fact, results from

Section 7.4 and Section 7.6 imply that short comings in conflict-resolution algorithms can

be overcome through better measurement and prediction of current and future aircraft po-

sitions. What makes such a result particularly interesting is that the computation times

required to improve trajectory predictions and position measurements are relatively small

when compared to improvements in conflict-resolution algorithms. However, from a in-

frastructure perspective, the cost of improved conflict-resolution tools depends on software

improvements coupled with more processing power. In contrast, to significantly improve the

position measurement of aircraft through ADS-B, related infrastructure costs are estimated

to be upwards of $7 billion [1], which excludes the development of software systems to take

advantage of the improvements.

For the remainder of the parameters and settings, the sensitivity analysis indicates

there is some leeway in their selection and implementation, assuming a reduction in aircraft

spacing. One of the more prominent results in this chapter indicates that the underlying

decision policy by which aircraft are selected to maneuver plays little role in determining

the conflict-resolution taskload over long periods of time. And as a corollary, the decision-

horizon time, HD , has little perceivable influence on the conflict-resolution taskload. Even

at high traffic intensities, the short-term benefits of the MCRT policy are not significant

until the 3X traffic intensity for the busiest sector in Minneapolis center.

Because selection of the decision policy is somewhat flexible, consideration of other

properties might be more relevant than just the conflict-resolution taskload. In an advisory

based setting, the air traffic controller is required to approve the safety of each resolution

command. It then might be beneficial for advisory systems to provide clear resolution com-

mands, that is, commands that can be readily identified as safe by the air traffic controller.

Doing so may allow the air traffic controller to continue processing potential conflicts in a

more efficient and safe manner, as less time is spent reviewing each advisory command.

148
 Also of interest, is the result that conflict-resolution algorithms need not provide conflict-

free travel for long periods of time. In fact, beyond guaranteed conflict-free times of HR = 6

minutes, there is limited benefit to improving conflict-resolution algorithms. An extension

to this result is that for some traffic intensities, it is better to improve solve-times to help

manage uncertainty than it is to increase conflict-free times. However, the division in trade-

off between conflict-free times and solve-times is case dependent on the traffic intensity and

traffic pattern within a sector.

Given that only a limited set of requirements is needed from a conflict-resolution algo-

rithm (i.e. a satisfactory conflict-free time HR , and a rapid solve-time δt), it is perhaps

within the best interest of air navigation service providers to push the development and

completion of a single advisory algorithm. Ultimately, the inclusion of an automated tool

will require significant testing and verification; in order to prevent delays, selection of a

single algorithm should be considered an imperative and occur as soon as possible.

149
 CHAPTER VIII

CONCLUSION

8.1 Summary

Historically, engineers, researchers, and designers of conflict-resolution algorithms for air

traffic systems have largely and implicitly ignored human factors issues by designing al-

gorithms to replace, rather than support air traffic controllers. Despite the advances in

automated conflict-resolution systems, there still exist a number of issues that are unad-

dressed or unsolved. In particular, current formulations are unable to guarantee safe and

dynamically feasible trajectory solutions.

The research presented in this thesis focuses on an alternative to automating tactical air

traffic control: the inclusion of advisory conflict-detection and resolution decision-support

tools to aid air traffic controllers, without replacing them. In a human-in-the-loop frame-

work, advisory conflict-detection and resolution systems identify potential conflicts and

then propose resolution commands for the air traffic controller to verify and issue to air-

craft. Much like automated or highly computerized solutions, advisory conflict-detection

and resolution tools can be designed to handle large traffic loads and provide solutions in

real-time. Since the air traffic controller is still part of the decision process and has the

option to accept or reject proposed solutions, there remains a safety fall-back.

While human factors and cognitive engineering researchers have highlighted key aspects

and requirements for the successful design of decision-support tools, there has been little

actualization of these concepts into mathematically rigorous algorithms. The major contri-

bution of this study is to bring together more formally these fields through the introduction

and analysis of a model to better understand the conflict-resolution taskload associated with

an advisory decision-support tool. The research presented here seeks to understand how

the formulation, capabilities, and implementation of conflict-detection and resolution tools

affect the controller taskload associated with the conflict-resolution process, and implicitly

150
controller workload.

Based on an analysis of a data-driven traffic model, and a sensitivity study of advisory

conflict-detection and resolution tools, the following insights are drawn:

• When traffic throughput increases, not only will the number of potential conflicts multiply,

but so will the complexity of conflict scenarios. Simple pairwise conflicts between two

aircraft will become less likely. Instead, conflict scenarios will contain multiple aircraft that

conflict with each other. (refer to Section 3.2)

• Without change to policies, practices, and technologies, air transportation demand will

one day exceed the capacity of the unaided air traffic controller. As such, there is a need

to provide some level of automation in air traffic control. (refer to Section 3.3)

• Currently, no guaranteed safe conflict-resolution algorithms exist. Therefore, an air traffic

system that relies on complete automation or supervisory control is not a tenable option.

Instead, advisory control might provide an acceptable alternative framework. Using advi-

sory systems, air traffic controllers are able to leverage the strength of automated systems,

while providing a safety fall back in the case of failure.

• The conflict-resolution taskload of an advisory conflict-detection and resolution system

is most sensitive to aircraft spacing requirements. To achieve the greatest reductions in

conflict-resolution taskload, air navigation service providers must make strong efforts to-

wards improving aircraft position measurements and trajectory prediction tools. One avenue

to achieving these goals is to make ADS-B utilization ubiquitous. (refer to Section 7.2 and

Section 7.4)

• For policies considered, studies demonstrate that the decision policy of a conflict-resolution

algorithm for selecting which aircraft to maneuver has an imperceivable effect on the conflict-

resolution taskload. Because of this insensitivity, there is leeway in the design of conflict-

resolution algorithms. So instead of spending effort to select a decision policy that minimizes

conflict-resolution taskload, research efforts can best be placed into other areas of interest

(e.g. minimizing fuel-costs). (refer to Section 7.3)

• Once aircraft spacing distances are reduced, the introduction of any conflict-resolution

algorithm that is capable of generating conflict-free trajectories for 6 minutes or longer

151
is satisfactory in managing conflict-resolution taskload (refer to Section 7.5). The lack of

strict requirements for an advisory conflict-detection and resolution system suggests that

any well specified and functioning algorithm can be used. To prevent delay in formulating

a complete tool, air navigation service providers and engineers should continue diligently

by proposing and testing other requirements needed for a conflict-resolution tool (e.g. fuel-

optimal, human-centric).

• No standard best practice exists for selecting the conflict-free resolution time and solve-

time. Instead, the selection of each depends on the prevailing traffic pattern within an

airspace and the traffic intensity. (refer to Section 7.6)

8.2 Contributions

According to the models and studies provided, the following contributions are contained

within this thesis:

• Based on generated traffic scenarios, a quadratic model that considers traffic intensity and

separation requirements is proposed to predict the number of uncontrolled conflicts within

an airspace. Similar to previous studies, the predicted number of uncontrolled conflicts is

proportional to the square of the traffic intensity. The major contribution of the quadratic

model is to indicate that the number of uncontrolled conflicts is proportional to aircraft

separation requirements. (refer to Section 3.2)

• Leveraging previous research representing aircraft and potential conflicts through static

conflict graphs, this thesis extends the graph model to the dynamic case. Furthermore this

thesis includes modeling of the conflict-detection and resolution process acting on dynamic

conflict graphs. (refer to Chapter 4)

• The term conflict-resolution taskload is coined. In regards to previous studies, the meaning

of conflict-resolution taskload is distinguished from the number of potential conflicts a

conflict-resolution tool solves. In this way, the conflict-resolution taskload is related to the

effort an air traffic controller makes in separating aircraft. (refer to Chapter 4)

• To support conflict-resolution taskload analysis of advisory decision-support tools, an

abstraction of the behaviors and characteristics of conflict-detection and resolution tools

152
is proposed. More specifically, the abstraction of conflict-detection tools considers how

potential conflicts are detected and resolved. Conflict-detection tools are parameterized

by aircraft spacing requirements and uncertainty in trajectory prediction. The abstraction

and parameterization of conflict-resolution tools considers the decision policy used to select

which aircraft are maneuvered; how much information is used in the decision making process;

the length of time for which resolution trajectories are guaranteed conflict-free; and how

often the conflict-resolution problem is solved. In addition to the abstraction of advisory

tools, a major contribution of this thesis is to provide a simulation framework in which to

implement them without specifying their exact behavior. (refer to Chapter 5)

• Application of the abstracted conflict-detection and resolution tools demonstrate that

for a controlled airspace, the conflict-resolution taskload is most sensitive to the traffic

intensity and the effective aircraft spacing. According to specified abstracted parameters

for a conflict-detection and resolution tool, the proposed model indicates that the conflict-

resolution taskload is proportional to I 2 Dsep

min , where I and D min refer to the traffic intensity
sep

and aircraft spacing. (refer to Section 7.1)

• Demonstration that improvements to conflict-detection systems would provide the greatest

benefit in reducing conflict-resolution taskload. (refer to Section 7.2 and Section 7.4)

• Based on a sensitivity analysis, a series of best and satisfactory practices is noted. One

specific result concludes that the decision policy for selecting which aircraft to maneuver

has little perceivable influence on the conflict-resolution taskload of an advisory conflict-

resolution system (refer to Section 7.3). Additionally, the thesis demonstrates that in-

creasing the conflict-free time of a conflict-resolution algorithm has diminishing returns for

reducing the conflict-resolution taskload (refer to Section 7.5).

8.3 Future Work

It is necessary to note that the research presented in this thesis does not consider all aspects

vital to determining conflict-resolution taskload. Furthermore, the results contained in

this thesis suggest additional avenues of research that will provide greater insight into the

interaction between air traffic controllers and advisory decision-support tools.

153
Improving Models. There is the potential for future work to incorporate a more flexi-

ble and realistic representation of the conflict-detection and resolution process. First, the

conflict-detection model can be expanded to consider non-isotropic growth in uncertainty as

well as variable growth rates. Such a model would be more reflective of errors in radar sys-

tems and trajectory prediction tools. Furthermore, to be more consistent with engineered

conflict-detection and resolution algorithms, potential conflicts and resolution commands

should be considered to be probabilistic. That is, potential conflicts are identified by their

probability of occurrence, and conflict-free times for resolution trajectories are stochastic.

System-wide Conflict Resolution. Future research should address the system-wide ef-

fects of advisory conflict-resolution systems. Specifically, how managing a single airspace

might result in downstream conflicts in other sectors. Or conversely, how conflict-resolution

tools can be designed to resolve potential conflicts not only within one sector, but future

sectors. The problem of considering adjacent sectors in the conflict-resolution process is

a complex task, however most likely needed. As traffic intensities increase, aircraft in-

teractions will become increasingly coupled through potential conflicts, not only at the

sector-level but at the center-level.

In this thesis, analysis of the conflict-free time, HR , relied on two assumptions. The

traffic generation procedure assumed that the aircraft trajectories and potential conflicts

generated for each airspace are independent of prior control actions taken in other sectors.

Furthermore, when generating resolution commands, the conflict-resolution model assumes

that solution trajectories are only conflict-free for HR minutes within the current sector.

Thus, the research as implemented in this thesis does not consider the multi-sector problem.

If a multi-sector approach towards conflict-resolution is desired then a serious discus-

sion of workload issues is needed. Any resolution command that adjusts trajectories in

future airspaces requires acceptance by the corresponding air traffic controller. In such

a framework, controllers do not act as independent units. They are required to propose

and approve trajectories with other controllers. While a system-wide approach towards

conflict-resolution might further reduce conflict-resolution taskload, there is uncertainty as

154
to how it will affect the workload of controllers. To enable such a system-wide approach

requires an information distribution system similar to Nextgen’s System Wide Information

management (SWIM) concept [50].

Guarantee conflict-free travel. Developing algorithms that function at high traffic

levels is still a concern. Even at present traffic levels there do not exist systems that can

guarantee conflict-free travel. Future research is required to strength conflict-resolution

algorithms.

Human-Centric Advisory Commands. The analysis in this thesis strictly focused on

the conflict-resolution taskload of decision-support tools. This approach was taken with

the understanding that it is a significant driver of controller workload when working with

trajectory based operations. Further, this thesis makes the assumption that all advisory

resolution commands affect controller workload equally. In fact, this assumption is probably

not true, especially during high traffic volumes or when advisory resolution trajectories are

difficult to decipher. Because the responsibility of ensuring safety lies with the air traffic

controller, the more complex an advisory solution is, the more effort a controller must make

in ensuring its safety and feasibility. So while a set of potential conflicts might be resolvable

with a small number of maneuvers, the complexity of advisory trajectories might require

significant mental effort from an air traffic controller. In these cases it might be preferred

to issue a sub-optimal number of resolution commands that the air traffic controller can

easily verify to be safe and feasible. Future studies are required to better understand how

controllers respond to complex trajectories.

There is a need to considered the ramifications of mixed aircraft equipage in regards

to ADS-B. Two of the underlying assumptions found in this thesis are that aircraft are

spaced uniformly, and that aircraft are equally capable of receiving resolution trajectories

consistent with trajectory based operations. For scenarios when ADS-B is not universal,

complications may arise in the conflict-resolution process. As such, additional studies on

the design of conflict-resolution algorithms in mixed-equipage environments is required.

Furthermore, the conflict-detection and resolution process proposed in this thesis does

155
 < HR
Resolution
maneuvers

(a) Current routes (b) Future routes

Figure 85: Taking full advantage of conflict-resolution systems may require a system-wide
airspace redesign.

not need to be conducted in a receding-horizon control framework constrained to fixed time-

steps and horizon times. Based on traffic conditions, the horizon times and the conflict-free

times can be adjusted to meet the needs of the controller. In control theory there is already

a stream of research that considers adaptive receding-horizon control [27, 65].

Airspace Redesign for Automated Conflict-Resolution. Finally, there is a need to

take a fundamentally different approach to understanding the implementation of advisory

tools. This thesis focused on analyzing the control system (both sensing and control) in

an effort to manage conflict-resolution taskload. An alternative approach is to redesign the

plant, e.g. the airspace, to be more consistent with advisory conflict-detection and resolution

systems. In current-day operations, major traffic flows are spatially separated to space the

time between potential conflicts for a single aircraft. Large distances between intersections

provide the air traffic controller with sufficient time between potential conflicts to generate

any necessary resolution commands. With trajectory based operations, if conflict-resolution

algorithms are able to guarantee conflict-free flight for HR minutes, then to take advantage

of advisory conflict-detection and resolution systems, sequential conflict regions should be

within HR minutes of each other. As illustrated in Figure 85, to reduce the number of

resolution commands, closely spaced flows might be beneficial. Otherwise, if intersection

are spaced far apart, aircraft might be required to re-plan trajectories each time a potential

conflict is identified.

156
 APPENDIX A

DECISION POLICIES

In addition to the decision policies detailed in Chapter 5, five other decision policies are

considered in the simulation analysis. They are detailed below.

A.1 First-Come, First-Served without Conflict-Priority Ordering

The First-Come, First-Served (FCFS) policy is adjusted to remove conflict-priority ordering.

The FCFS policy without conflict-priority ordering results in a slightly different process for

resolving potential conflicts when compared to the original policy described in Procedure 2.

Instead of processing and resolving potential conflicts according to the order they occur,

aircraft are first sorted by their arrival time into the airspace. The most recent unresolved

aircraft involved in potential conflict is selected to receive a resolution command. The

temporary edge-set E T is updated accordingly. The processes repeats until the edge set E T

is empty. The final set of maneuvered aircraft is given by the set M. The complete process

is provided in Procedure 6.

157
 Procedure 6 First-Come, First-Served with-

out Conflict-Priority Ordering

E T ← Ê(k), q = 0

while card(E T ) do

q =q+1

i∗ = argmax(tai | (ni , ∗) ∈ E T )
i
mTq = n i∗

E T ← E T \ (mTq , ∗)

end while

I = [t + TR , t + TR + δt]

M = {m | m ∈ mT ∩ ∃tcm,∗ ∈ I}

A.2 First-Exit, First-Served

The first-exit, first-served policy (FEFS) is similar to the FCFS policy; however, instead of

prioritizing aircraft based on their arrival time into the sector, aircraft are given priority

according to when they are scheduled to exit the sector.

Using the same notation found in the procedure for the FCFS policy (see Procedure 2

and Procedure 6), the procedures for FEFS, with and without conflict-priority ordering,

are given by Procedure 7 and Procedure 8. In the case of the FEFS policy, aircraft are

ordered according to the original scheduled exit time from the sector, txi . In the case when

an aircraft receives a resolution maneuver, potentially adjusting the expected exit time, the

scheduled exit time remains the same.

158
Procedure 7 First-Exit, First-Served with Procedure 8 First-Exit, First-Served with-

Conflict-Priority Ordering out Conflict-Priority Ordering

E T ← Ê(k), q = 0 E T ← Ê(k), q = 0

while card(E T ) do while card(E T ) do

q =q+1 q =q+1

(i∗ , j ∗ ) = argmin tci,j i∗ = argmin(txi | (ni , ∗) ∈ E T )

i,j i
mTq = n i∗

ni∗ if txj∗ < txi∗


T
mq = E T ← E T \ (mTq , ∗)

n ∗ else

j
end while
ET ← ET \ (mTq , ∗)
I = [t + TR , t + TR + δt]
end while
M = {m | m ∈ mT ∩ ∃tcm,∗ ∈ I}
I = [t + TR , t + TR + δt]

M = {m | m ∈ mT ∩ ∃tcm,∗ ∈ I}

Like the FCFS policy, there is always the potential of favoring aircraft traveling along

specific routes. However, a benefit of the procedure is that because aircraft closer to sector

boundaries are prioritized, the advisory system is less likely to issue resolution commands to

aircraft about to transitioning between sectors. From the human-factors perspective, such

a property may be preferred, as maneuvering aircraft near boundaries requires additional

coordination effort between the controllers in each sector.

A.3 Knowledge-Based Policy

This policy is based on a knowledge-based policy called Resolution Aircraft and Maneuver

Selector (RMAS). RAMS forms the backbone of the Automated Airspace Concept [32], a

tool designed for performing automated conflict-resolution. Based on the characteristics

of a pairwise conflict, RAMS selects which of the two aircraft to maneuver, as well as

the preferred resolution command, according to a rule-set. RAMS is based on standard

practices derived from human controllers and operational insights and analytical studies.

In such a framework, potential conflicts are resolved according to conflict-priority ordering.

159
 Departure Ascent Cruise Descent Arrival

Airport Airport

Figure 86: Different phases of flight considered in the knowledge-based policy.

The RAMS’ rules dictating the preferred aircraft to maneuver for each type of potential

conflict are similar to those provided in Table 9. In the original formulation of RAMS, an

ordering of preferred solutions is included as part of the policy. Each potential conflict is

defined according to the phase of flight each aircraft is in at the time of occurrence, not

the time of detection. The major phases of flight include: ascent, cruise, and descent.

The flight phases are illustrated in Figure 86. Table 9 provides a simplification of the rules

given in [32]. As part of the RAMS system, additional fidelity differentiated two types of

arrivals: cruise arrivals and descending arrivals. For the RAMS system, an arrival aircraft

is defined as any aircraft within 200 NM or 20 minutes of its TRACON arrival fix. Because

information concerning the distance and time to TRACON arrival fixes is not available in

the PDARS data used to generate the traffic simulations, some of the rules are collapsed

together to still allow for an approximation of RAMS. In some cases, the simplification loses

resolution procedures that prevent reconstruction or inversion of the original rules.

Table 9: Knowledge-Based Policy Rules

Flight Phase of Aircraft

Aircraft 1 Aircraft 2 Preferred Aircraft to maneuver
Cruise Cruise Aircraft furthest from airspace
Large crossing angle boundary or top of descent
Cruise Cruise
Faster aircraft
Small crossing angle or aircraft in trail
Cruise Ascent Ascending aircraft
Cruise Descent Cruising aircraft
Ascent Ascent Lower aircraft
Ascent Descent Ascending aircraft
Descent Descent Trailing aircraft

The procedure for the knowledge-based policy with conflict-priority ordering is provided

in Procedure 9. The function KBR(Ai , Aj ), included as part of the policy procedure, enacts

160
the knowledge-based rules provided in Table 9. The input declaration of the aircraft pair

(Ai , Aj ) into the function KBR(Ai , Aj ) assumes that classification information is available.

Procedure 9 Knowledge-Based Policy with

Conflict-Priority Ordering
E T ← Ê(k), q = 0

while card(E T ) do

q =q+1

(i∗ , j ∗ ) = argmin tci,j

i,j
mTq = KBR(Ai∗ , Aj ∗ )

E T ← E T \ (mTq , ∗)

end while

I = [t + TR , t + TR + δt]

M = {m | m ∈ mT ∩ ∃tcm,∗ ∈ I}

A.4 Random Policy with Maneuver Balancing

The second random procedure, described by Procedure 10, takes a more balanced approach

towards deciding which aircraft are issued resolution commands. In particular, the policy

considers which aircraft have been previously maneuvered through the vector M P . The

value of MiP indicates the number of times aircraft Ai has been issued a resolution command.

For the potential conflict between aircraft Ai and Aj , the values MiP and MjP are compared.

If MiP = MjP , that is, both aircraft have been maneuvered the same number of times, then

the policy randomly selects which aircraft to maneuver. In the case when MiP 6= MjP ,

the aircraft with less maneuvers is issued a resolution command. After the corresponding

aircraft are issued resolution commands, the vector M P is updated.

By taking into account the number of times an aircraft has been maneuvered prior to any

potential conflicts, the second policy acts to balance the number of times any one aircraft

receives trajectory perturbations. In the context of fuel consumption, if each maneuver

161
entails an extra deviation or fuel-cost, then by balancing the distribution of maneuvered

aircraft, the fuel consumption is better balanced across aircraft.

Procedure 10 Random with Prior-

Resolution Balancing and Conflict-Priority

Ordering
E T ← hatE(k), q = 0

while card(E T ) do

q =q+1

(i∗ , j ∗ ) = argmin tci,j

i,j

ni∗ if (MiP∗ < MjP∗ ) ∪







T
mq = (MiP∗ = MjP∗ ∩ rand < .5)





n ∗ if else

j


E T ← E T \ (mTq , ∗)

end while

I = [t + TR , t + TR + δt]

M = {m | m ∈ mT ∩ ∃tcm,∗ ∈ I}

MiP ← MiP + card(i ∈ M)

162
 REFERENCES

[1] “Final regulatory impact analysis, final rule, Automatic Dependent Surveillance
Broadcast (ADS-B) Out, performance requirements to support air traffic control ser-
vice,” Tech. Rep. FAA200729305, Federal Aviation Administration, Office of Aviation
Policy and Plans, August 18, 2009.

[2] Agogino, A. and Tumer, K., “Evolving distributed agents for managing air traffic,”
in Conference on Genetic and Evolutionary Computation.

[3] Alliot, J., Bosc, J., Durand, N., and Maugis, L., “CATS: A complete air traffic
simulator,” in Digital Avionics Systems Conference, 1997.

[4] Archambault, N. and Durand, N., “Scheduling heuristics for on-board sequential
air conflict solving,” in Digital Avionics Systems Conference, 2004.

[5] Arthur, W. and McLaughlin, M., “User request evaluation tool (URET)
interfacility conflict probe performance assessment,” in USA/Europe Air Traffic
Management R&D Seminar, 1998.

[6] Averty, P., “Conflict Perception by ATCS Admits Doubt but not Inconsistency,”
in USA/Europe Air Traffic Management R&D Seminar, 2005.

[7] Averty, P., “Elements for prioritizing between conflict resolutions in air traffic con-
trol,” in Digital Avionics Systems Conference, 2008.

[8] Barnier, N. and Allignol, C., “4D-trajectory deconfliction through departure

time adjustment,” in USA/Europe Air Traffic Management R&D Seminar, 2009.

[9] Billings, C. and Woods, D., “Concerns about adaptive automation in aviation
systems,” Human performance in automated systems: Current research and trends,
pp. 264–269, 1994.

[10] Bilmoria, K., Banavar, S., Chatterji, G., Sheth, K., and Grabbe, S.,
“FACET: Future ATM concepts evaluation tool,” Air Traffic Control Quarterly, vol. 9,
no. 1, 2000.

[11] Brooker, P., “Control workload, airspace capacity and future systems,” Human
Factors and Aerospace Safety, vol. 3, no. 1, pp. 1–23, 2003.

[12] Calderón-Meza, G. and Sherry, L., “Analysis of stakeholder benefits of

nextgen trajectory-based operations,” in Integrated Communications Navigation and
Surveillance Conference, 2010.

[13] Calderón-Meza, G. and Sherry, L., “Establishing an Upper-Bound for the

Benefits of NextGen Trajectory-Based Operations,” in International Conference on
Research in Air Transportation, 2010.

163
[14] Cardosi, K., “An Analysis of En Route Controller-Pilot Voice Communications,”
Tech. Rep. DOT/FAA/RD-93/11, US Department of Transportation, Federal Avia-
tion Administration, 1993.

[15] Chaloulos, G., Hokayem, P., and Lygeros, J., “Distributed hierarchical MPC
for conflict resolution in air traffic control,” in American Control Conference.

[16] Chaloulos, G., Roussos, G., Lygeros, J., and Kyriakopoulos, K., “Ground
Assisted Conflict Resolution in Self-Separation Airspace,” in Guidance, Navigation
and Control Conference and Exhibit, 2008.

[17] Chatterji, G. and Sridhar, B., “Neural network based air traffic controller work-
load prediction,” in American Control Conference, 1999.

[18] Chvatal, V., “A greedy heuristic for the set-covering problem,” Mathematics of
Operations Research, pp. 233–235, 1979.

[19] Clarkson, K., “Modification of the greedy algorithm for vertex cover.,” Information
Processing Letters, vol. 16, no. 1, pp. 23–26, 1983.

[20] Committee for a Review of the En Route Air Traffic Control Com-
plexity and Workload Model, “Air Traffic Controller Staffing in the En Route
Domain : A Review of the Federal Aviation Administrations Task Load Model,” tech.
rep., National Research Council of the National Academies, 2010.

[21] Corker, K., Gore, B., Fleming, K., and Lane, J., “Free Flight and the Con-
text of Control: Experiments and Modeling to Determine the Impact of Distributed
Air-ground Air Traffic Management on Safety and Procedures,” in USA/Europe Air
Traffic Management R&D Seminar, 2000.

[22] Cruck, E. and Lygeros, J., “Subliminal Air Traffic Control: Human Friendly
Control of a Multi-Agent System,” in American Control Conference, 2007.

[23] Davis, T., Krzeczowski, K., and Bergh, C., “The final approach spacing tool,”
in Symposium on Automatic Control in Aerospace, 1994.

[24] Delahaye, D. and Puechmorel, S., “Air traffic complexity map based on non lin-
ear dynamical systems,” in EUROCONTROL Innovative Research Workshop, 2005.

[25] Dougui, N., Delahaye, D., Puechmorel, S., and Mongeau, M., “A New
Method for Generating Optimal Conflict Free 4D Trajectory,” in International
Conference on Research in Air Transportation, 2010.

[26] Dowek, G., Muñoz, C., and Carreño, V., “Provably Safe Coordinated Strategy
for Distributed Conflict Resolution,” in Guidance Navigation, and Control Conference
and Exhibit, 2005.

[27] Droge, G. and Egerstedt, M., “Adaptive look-ahead for robotic navigation in un-
known environments,” in International Conference on Intelligent Robots and Systems,
2011.

[28] Durand, N. and Alliot, J.-M., “Optimal resolution of en route conflicts,” in

USA/Europe Air Traffic Management R&D Seminar, 2001.

164
[29] Ehrmanntraut, R., “The Potential of Speed Control,” in Digital Avionics Systems
Conference, 2004.

[30] Endsley, M., “Automation and situation awareness,” Automation and human
performance: Theory and applications, pp. 163–181, 1996.

[31] Ennis, R. and Zhao, Y., “A formal approach to the analysis of aircraft protected
zone.,” Air Traffic Control Quarterly, vol. 12, no. 1, pp. 75–102, 2004.

[32] Erzberger, H., “Automated Conflict Resolution for Air Traffic Control,” in
International Congress of the Aeronautical Sciences, 2006.

[33] Erzberger, H. and Paielli, R., “Concept for next generation air traffic control
system.,” Air Traffic Control Quarterly, vol. 10, no. 4, pp. 355–378, 2002.

[34] Erzberger, H., Paielli, R., Isaacson, D., and Eshow, M., “Conflict detec-
tion and resolution in the presence of prediction error,” in USA/Europe Air Traffic
Management R&D Seminar, 1997.

[35] Federal Aviation Administration, “FAA Aerospace Forecast: Fiscal Years 2010-
2030,” tech. rep., U.S. Department of Transportation, 2010.

[36] Federal Aviation Administration, “Facility Orientation Guide: Minneapolis

ARTCC.” www.stuckmic.com/faa-facility-orientation-guides.html, September 2011.

[37] Galdino, A., Muñoz, C., and Ayala, M., “Formal Verification of an Optimal
Air Traffic Conflict Resolution and Recovery Algorithm,” in Workshop on Logic,
Language, Information and Computation, 2007.

[38] Gangel, M., Mehta, G., Muhammad, C., Sekhavat, S., Vora, G., and Arkind,
K., “Terminal area capacity enhancement concept,” in Systems and Information
Engineering Design Symposium, 2004.

[39] Garcia, J., Besada Portas, J., Soto Jaramillo, A., and Miguel Vela, G.,
“Opportunity trajectory reconstruction techniques for evaluation of ATC systems,”
International Journal of Microwave and Wireless Technologies, vol. 1, no. 3, pp. 231–
238, 2009.

[40] Gawinowski, G., Garcia, J.-L., Guerreau, R., Weber, R., and Brochard,
M., “ERASMUS: A new Path for 4D Trajectory-Based Enablers to Reduce the Traffic
Complexity,” in Digital Avionics Systems Conference, 2007.

[41] Granger, G. and Durand, N., “A traffic complexity approach through cluster
analysis,” in USA/Europe Air Traffic Management R&D Seminar, 2003.

[42] Green, S., Vivona, R., Grace, M., and Fang, T., “Field evaluation of descent
advisor trajectory prediction accuracy for en route clearance advisories,” in Guidance,
Navigation, and Control Conference and Exhibit, 1998.

[43] Harding, S. and Banzhaf, W., “Fast genetic programming on GPUs,” Genetic
Programming, pp. 90–101, 2007.

165
[44] Hinton, D., Charnock, J., and Bagwell, D., “Design of an aircraft vortex
spacing system for airport capacity improvement,” in Aerospace Sciences Meeting
& Exhibit, 2000.

[45] Hoekstra, J., van Gent, R., and Ruigrok, R., “Designing for safety: the free
flight air traffic management concept,” Reliability Engineering & System Safety,
vol. 75, no. 2, pp. 215–232, 2002.

[46] Hull, J., Barmore, B., and Abbott, T., “Technology-enabled airborne spacing
and merging,” in Digital Avionics Systems Conference, 2004.

[47] Irvine, R., “A simplified approach to conflict probability estimation,” in Digital

Avionics Systems Conference, 2001.

[48] Jardin, M., “Air traffic conflict models,” in Aviation Technology, Integration and
Operations Conference, 2004.

[49] Joint Planning and Development Office, “Next Generation Air Transportation
System Integrated Work Plan: A Functional Outline,” tech. rep., Joint Planning and
Development Office, Washington, DC, Sept. 2008.

[50] Joint Planning and Development Office, “Concept of Operations for the Next
Generation Air Transportation System,” tech. rep., Joint Planning and Development
Office, Washington, DC, Sept. 2010.

[51] Kirwan, B., “The role of the controller in the accelerating industry of air traffic
management,” Safety Science, vol. 37, no. 2-3, pp. 151–185, 2001.

[52] Kirwan, B. and Flynn, M., “Towards a controller-based conflict resolution tool -
a literature review,” Tech. Rep. ASA.01.CORA.2.DEL04-A.LIT, EUROCONTROL,
2002.

[53] Kopardekar, P., Prevot, T., and Jastrzebski, M., “Traffic complexity mea-
surement under higher levels of automation and higher traffic densities,” Air Traffic
Control Quarterly, vol. 17, no. 2, pp. 125–148, 2009.

[54] Kopardekar, P., Schwartz, A., Magyarits, S., and Rhodes, J., “Airspace
complexity measurement: An air traffic control simulation analysis,” in USA/Europe
Air Traffic Management R&D Seminar, 2007.

[55] Kopardekar, P. and Magyarits, S., “Dynamic density: Measuring and predicting
sector complexity,” in Digital Avionics Systems Conference, 2002.

[56] Kuchar, J. and Yang, L., “A Review of Conflict Detection and Resolution Modeling
Methods,” IEEE Transactions on Intelligent Transportation Systems, vol. 1, pp. 179–
189, Dec. 2000.

[57] Lange, M., Hjalmarsson, J., Cooper, M., Ynnerman, A., and Duong, V., “3d
visualization and 3d and voice interaction in air traffic management,” in The annual
SIGRAD Conference, 2003.

[58] Laudeman, I., Shelden, S., Branstrom, R., and Brasil, C., “Dynamic density:
An air traffic management metric,” Tech. Rep. NASA/TM-1998-112226, 1998.

166
[59] Loft, S., Sanderson, P., Neal, A., and Mooij, M., “Modeling and predicting
mental workload in en route air traffic control: Critical review and broader implica-
tions,” Human Factors: The Journal of the Human Factors and Ergonomics Society,
vol. 49, no. 3, p. 376, 2007.
[60] Loghides, C., Paulson, G., and Philipp, W., “Guidelines for the application
of the ecac radar separation minima,” Tech. Rep. ASM.ET1.ST18.1000 -REP-01.00,
EUROCONTROL, 1998.
[61] Maddalon, J., Butler, R., Geser, A., and Muñoz, C., “Formal verification
of a conflict resolution and recovery algorithm,” Tech. Rep. NASA/TP-2004-213015,
NASA Langley Research Center, NASA LaRC,Hampton VA 23681-2199, USA, April
2004.
[62] Manning, C., Mills, S., Pfleiderer, E., Fox, C., and Mogilka, H., “Relation-
ships among observer ratings of mental workload and taskload measures derived from
routinely recorded air traffic control data,” Focusing Attention on Aviation Safety,
2001.
[63] Manning, C., Mills, S., Fox, C., Pfleiderer, E., and Mogilka, H., “Using Air
Traffic Control Taskload Measure and Communication Events to Predict Subjective
Workload,” tech. rep., Federal Aviation Administration, Office of Aerospace Medicine,
2002.
[64] Metzger, U. and Parasuraman, R., “The role of the air traffic controller in future
air traffic management: An empirical study of active control versus passive monitor-
ing,” Human Factors: The Journal of the Human Factors and Ergonomics Society,
vol. 43, no. 4, pp. 519–528, 2001.
[65] Michalska, H. and Mayne, D., “Robust receding horizon control of constrained
nonlinear systems,” IEEE Transactions on Automatic Control, vol. 38, no. 11,
pp. 1623–1633, 1993.
[66] Mogford, R., “The complexity construct in air traffic control: a review and synthesis
of the literature,” Tech. Rep. DOT/FAA/CT-TN92/22, Department of Transporta-
tion / Federal Aviation Administration Technical Center, 1995.
[67] Mondoloni, S., Paglione, M., and Green, S., “Trajectory modeling accuracy for
air traffic management decision support tools,” in Congress of International Council
of the Aeronautical Sciences, 2002.
[68] Mueller, E. and Lozito, S., “Experimental evaluation of an integrated
datalink and automation-based strategic trajectory concept,” in Aviation Technology,
Integration, and Operations Conference, 2007.
[69] Niedringhaus, W., “Maneuver Option Manager: Automated Simplification of
Complex Air Traffic Control Problems,” IEEE Transactions on Systems, Man and
Cybernetics, vol. 22, no. 5, pp. 1047–1057, 1991.
[70] Niedringhaus, W., “Stream Option Manager (SOM): Automated Integration of
Aircraft Separation, Merging, Stream Management, and other Air Traffic Control
Functions,” IEEE Transactions on Systems, Man and Cybernetics, vol. 25, no. 9,
pp. 1269–1280, 1995.

167
[71] Nijhuis, H., “RHEA, Role of the Human in the Evolution of ATM systems: Final
Report,” tech. rep., National Aerospace Laboratory, 1998.

[72] Ore, O., Theory of Graphs. American Mathematical Society, 1962.

[73] Ozgur, M. and Cavcar, A., “A knowledge-based conflict resolution tool for rn-
route air traffic controllers,” Aircraft Engineering and Aerospace Technology, vol. 80,
no. 6, pp. 649–656, 2008.

[74] Paielli, R. and Erzberger, H., “Conflict Probability Estimation for Free Flight,”
Journal of Guidance, Control, and Dynamics, vol. 20, no. 3, pp. 588–596, 1997.

[75] Parasuraman, R. and Riley, V., “Humans and automation: Use, misuse, disuse,
abuse.,” Human Factors, vol. 39, no. 2, 1997.

[76] Perry, T., “In search of the future of air traffic control,” Spectrum, IEEE, vol. 34,
no. 8, pp. 18–35, 1997.

[77] Porterfield, D., “Evaluating controller communication time as a measure of work-

load,” International Journal of Aviation Psychology, vol. 7, no. 2, pp. 171–182, 1997.

[78] Prandini, M., Putta, V., and Hu, J., “A probabilistic measure of air traffic com-
plexity in three-dimensional airspace,” International Journal of Adaptive Control and
Signal Processing, vol. 1, pp. 1–25, 2009.

[79] Prevot, T., Lee, P., Callantine, T., Smith, N., and Palmer, E.,
“Trajectory-oriented time-based arrival operations: Results and recommendations,”
in USA/Europe Air Traffic Management R&D Seminar, 2003.

[80] Rardin, R. L., Optimization in Operations Research. Prentice Hall, 1997.

[81] Reynolds, T. and Hansman, R., “Analysis of separation minima using a surveil-
lance state vector approach,” in USA/Europe Air Traffic Management R&D Seminar,
2000.

[82] Ropella, K., Introduction to Statistics for Biomedical Engineers. Morgan and
Claypool Publishers, 2007.

[83] Roussos, G., Chaloulos, G., Kyriakopoulos, K., and Lygeros, J., “Control
of multiple non-holonomic air vehicles under wind uncertainty using model predic-
tive control and decentralized navigation functions,” in Conference on Decision and
Control.

[84] Salaün, E., Gariel, M., Vela, A., Feron, E., and Clarke, J., “Airspace Com-
plexity Estimations Based on Data-Driven Flow Modeling,” in Guidance, Navigation,
and Control Conference, 2010.

[85] Schmidt, D., “On the conflict frequency at air route intersections,” Transportation
Research, vol. 11, no. 5, pp. 351–355, 1977.

[86] SESAR Consortium, “The ATM Target Concept: D3,” Tech. Rep. DLM-0612-001-
02-00, SESAR Consortium, Toulouse, France, Sept. 2007.

168
 [87] Shingledecker, C. A., “Subsidiary radio communications tasks for workload assess-
ment in R&D simulations,” tech. rep., Air force aerospace medical research laboratory,
1982.
[88] Shyu, S., Yin, P., and Lin, B., “An ant colony optimization algorithm for the
minimum weight vertex cover problem,” Annals of Operations Research, vol. 131,
no. 1, pp. 283–304, 2004.
[89] Slattery, R. and Green, S., “Conflict free trajectory planning for air traffic control
automation,” Tech. Rep. TM-108790, NASA Ames Research Center, 1994.
[90] Smith, A. and Bateman, H., “Management of holding patterns: A potential ads-b
application,” in Digital Avionics Systems Conference, 2008.
[91] Smith, P., McCoy, C., and Layton, C., “Brittleness in the design of cooperative
problem-solving systems: The effects on user performance,” IEEE Transactions on
Systems, Man and Cybernetics, Part A: Systems and Humans, vol. 27, no. 3, pp. 360–
371, 2002.
[92] Sperandio, J., “Variation of Operator’s Strategies and Regulating Effects on Work-
load,” Ergonomics, vol. 14, no. 5, pp. 571–577, 1971.
[93] Sridhar, B., Sheth, K., and Grabbe, S., “Airspace complexity and its application
in air traffic management,” in USA/Europe Air Traffic Management R&D Seminar,
1998.
[94] Stein, E., “Human Operator Workload in Air Traffic Control,” Human Factors in
Air Traffic Control, pp. 155–184, 1998.
[95] Swenson, H., Barhydt, R., and Landis, M., “Next Generation Air Transporta-
tion System (NGATS) Air Traffic Management (ATM)-Airspace Project,” tech. rep.,
National Aeronautics and Space Administration, 2006.
[96] Swierstra, S. and Green, S., “Common trajectory prediction capability for deci-
sion support tools,” in USA/Europe Air Traffic Management R&D Seminar, 2003.
[97] Tomlin, C., Pappas, G., Lygeros, J., Godbole, D., and Sastry, S., “Hy-
brid control models of next generation air traffic management,” Hybrid systems IV,
pp. 378–404, 1997.
[98] Transforming the NAS: The next generation air traffic control system,
“Erzberger, h.,” in International Congress of the Aeronautical Sciences, 2005.
[99] Transportation Safety Board of Canada, “Aviation re-
ports - 2008 - a08w0007.” http://www.tsb.gc.ca/eng/rapports-
reports/aviation/2008/a08w0007/a08w0007.asp, 9 2011.
[100] Vela, A., Solak, S., Clarke, J., Singhose, W., Johnson, E., and Barnes, E.,
“Near Real-Time Fuel-Optimal En Route Conflict Resolution,” IEEE Transactions
on Intelligent Transportation Systems, vol. 11, no. 4, pp. 826–837, 2010.
[101] Vela, A., Solak, S., Feron, E., Feigh, K., Singhose, W., and Clarke, J., “A
fuel optimal and reduced controller workload optimization model for conflict resolu-
tion,” in Digital Avionics Systems Conference, 2009.

169
[102] Wang, H., Kirubarajan, T., and Bar-Shalom, Y., “Precision large scale air traf-
fic surveillance using imm/assignment estimators,” IEEE Transactions on Aerospace
and Electronic Systems, vol. 35, no. 1, pp. 255–266, 1999.

[103] Wanke, C., “Using Air-Ground Data Link to Improve Air Traffic Management De-
cision Support System Performance,” in USA-Europe ATM R&D Seminar, 1997.

[104] West, D. B., Introduction to Graph Theory (2nd Edition). Prentice Hall, Aug. 2000.

170