An Intelligent Approach to Coordinated Control of Multiple Unmanned Aerial Vehicles

George Vachtsevanos, Liang Tang, Johan Reimann

gjv@ece.gatech.edu ltang@ece.gatech.edu gtg221d@prism.gatech.edu
School of Electrical and Computer Engineering
Georgia Institute of Technology, Atlanta, GA, 30332. U.S.A.

Abstract

This paper introduces a novel architecture for the coordinated control of multiple Unmanned Aerial Vehicles (UAVs ) and a
differential game theoretical approach to formation control and collision avoidance. The hierarchical architecture features an
upper level with global situation awareness and team mission planning, a middle level with local knowledge, formation control
and obstacle avoidance, and a low level that interfaces with onboard baseline controllers, sensors, communication and weapon
systems. Each level consists of several interacting agents with dedicated functions. The formation control problem is viewed as a
Pursuit Game of n pursuers and n evaders. Stability of the formation of vehicles is guaranteed if the vehicles can reach their
destinations within a specified time, assuming that the destination points are avoiding the vehicles in an optimal fashion. A two-
vehicle example is shown to illustrate the approach. Vehicle model is simplified to point mass with acceleration limit. Collision
avoidance is achieved by designing the value function so that it ensures that the two vehicles move away from one another when
they come too close to each one. Simulation results are provided to verify the performance of the proposed algorithm.

In this paper, we introduce a novel architecture for the

Introduction
The future urban warfare will utilize an unprecedented level
of automation in which human-operated, autonomous, and
semi -autonomous air and ground platforms will be linked
through a coordinated control system. Networked UAVs
bring a new dimension to future combat systems that must
include adaptable operational procedures, planning and
deconfliction of assets coupled with the technology to
realize such concepts. The technical challenges the control
designer is facing for autonomous collaborative operations
stem from real-time sensing, computing and
communications requirements, environmental and
operational uncertainty, hostile threats and the emerging
need for improved UAV and UAV team autonomy and
reliability. Figure 1 shows the autonomous control level
trend according to the DoD UAV Roadmap (Ref. 1). To
meet these challenges, innovative coordinated planning and
control technologies such as distributed artificial
intelligence (DAI), computational intelligence and soft
computing, as well as game theory and dynamic
optimization, have been investigated intensively in recent
years. However, in this area, more work has been focused
on solving particular problems, such as formation control
and autonomous search, while less attention has been paid
to the system architecture, especially from an
implementation and integration point of view.
Presented at the American Helicopter Society 60th Annual
Forum, Baltimore, MD, June 7-10, 2004. Copyright © 2004
by the American He licopter Society International, Inc. All
rights reserved.
coordinated control of multiple UAVs acting as intelligent
agents. Our approach differs from most others (Ref. 2,3) in
that, 1). A commander is placed at the highest level of the
hierarchy. At the current level of autonomy, the system
under development is acting as a decision support tool for
the commander. 2) The architecture is generic and flexible
to facilitate the fusion of diverse technologies.
The hierarchical architecture we introduce features an upper
level with global situation awareness and team mission
planning, a middle level with local knowledge, formation
control and obstacle avoidance, and a low level that
interfaces with onboard baseline controller, sensors,
communication and weapon systems. Each level consists of
several interacting agents with dedicated functions. Among
all the functional agents accommodated by the proposed
architecture, this paper focuses on formation control and
collision avoidance.
The problem of finding a control algorithm, which will
ensure that multiple autonomous vehicles can maintain a
formation while traversing a desired path and avoid inter-
vehicle collisions, will be referred to as the formation
control problem. The formation control problem has
recently received considerable attention due in part to its
wide range of applications in aerospace and robotics. A
classic example involving the implementation of the virtual
potential problem is presented in (Ref. 4). The authors
performed simulations on a two-dimensional system, which
proved to be well behaved. However, as they mention in
their conclusion, the drawback of the virtual potential
function approach is the possibility of being “trapped” in
local minima. Hence, if local minima exist, one cannot
guarantee that the system is stable. In (Ref. 5), the
individual trajectories of autonomous vehicles moving in
formation were generated by solving the optimal control
problem at each time step. This is computationally
demanding and hence not possible to perform in real-time
with current hardware.
acceptable formation maneuvers that can be performed
while maintaining the formation.
System Architecture
While networked and autonomous UAVs can be centrally
controlled, this requires that each UAV communicates all
the data from its sensors to a central location and receives
all the control signals back. Network failures and
communication delays are one of the main concerns in the
design of cooperative control systems. On the other hand,
distributed intelligent agent systems provide an environment
in which agents autonomously coordinate, cooperate,
negotiate, make decisions and take actions to meet the
objectives of a particular application or mission. The
autonomous nature of agents allows for efficient
communication and processing among distributed resources.

For the purpose of coordinated control of multiple UAVs,

Figure 1. Autonomous Control Level Trend
This paper views the formation control problem from a two
player differential game perspective, which provides a
framework to determine acceptable initial vehicle
deployment conditions but also, provides insight into
each individual UAV in the team is considered as an agent
with particular capabilities engaged in executing a portion
of the mission. The primary task of a typical team of UAVs
is to execute faithfully and reliably a critical mission while
satisfying local survivability conditions. In order to define
the domain of our research, we adopt an assumed mission
scenario of a group of 5 UAVs executing reconnaissance
and surveillance (RS) missions in an urban warfare
environment, as depicted in Figure 2.

Urban Warfare
GTMax Manned Vehicle

Fixed Wing UAV GTMax

GTMav

OAV Sniper

Ground
Ground
Sensor Sensor

Moving
Target
Soldiers Ground
Sensor

Commander
Operator

Figure 2. A Team of 5 UAVs Executing RS Missions in an Urban Warfare Environment

 Command Manned
& Control Vehicle

Level 3
Global
Team Mission Planning Global Situation Knowledge Fusion
Knowledge
/Re-planning Agent Awareness Agent Agent
Global Performance QoS Assessment
Measurement Agent Agent

Level 2
Local Formation Control Moving Obstacle Local Situation
Knowledge Agent Avoidance Agent Awareness Agent

Local Mission FDI/Reconfigurable

Planning Agent Control Agent

Level 1
Behavioral
Vehicle Weapon System
Knowledge Communication Sensing Agent
Control Agent Agent Agent

……

Figure 3. A Generic Hierarchical Multi-agent System Architecture

Worth pointing out is that the team consists of
heterogeneous UAVs with various capabilities in terms of
maneuverability, sensing, endurance, autonomy level, etc.
This mission scenario covers a wide area of research
interests including mission planning and re-planning,
optimal sensor (UAV) placement, task allocation,
autonomous search and tracking, knowledge fusion,
formation control, moving obstacle avoidance, fault tolerant
control, etc. A generic hierarchical multi-agent architecture
that accommodates the aforementioned technologies is
depicted in Figure 3.
The highest level of the control hierarchy features functions
of global situation awareness and teamwork. The mission
planning agent is able to generate and refine mission plans
for the team, generate or select flight routes, and create
operational orders. It is also responsible for keeping track of
the team’s plan, goals, and team members' status. The
overall mission is usually planned by the command and
control center based on the capabilities of each individual
UAV agent, and is further decomposed into tasks/subtasks
which are finally allocated to the UAV assets (individually
or in coordination with other vehicles). This can usually be
cast as a constrained optimization problem and tackled with
various approaches, such as integer programming, graph
theory, etc. Market based methods (Ref. 6, 7), and especially
auction theory (Ref. 8, 9, 10) can be applied as a solution to
autonomous mission re-planning. Planning the UAVs’ flight
route is also an integral part of mission planning. A modified
A* search algorithm, which attempts to minimize a suitable
cost function consisting of the weighted sum of distance,
hazard and maneuverability measures (Ref. 11,12), can be
utilized to facilitate the design of the route planner. In the
case of a leader-follower scenario, an optimal route is
generated for the leader, while the followers fly in close
formation in the proximity of the leader. The global situation
awareness agent, interacting with the knowledge fusion
agent, evaluates the world conditions based on data gathered
from each UAV (and ground sensors if available) and
reasons about the enemy's likely actions. Adversarial
reasoning and deception reasoning are two important tasks
executed here. The global performance measurement agent
measures the performance of the team and suggests team re-
configuration or mission re-planning, whenever necessary.
Quality of service (QoS) is assessed to make the best effort
to accomplish the mission and meet the predefined quality
criteria. Real world implementation of this level is not
limited to the agents depicted in the figure. For example, in
heterogeneous agent societies, knowledge of coordination
protocols and languages may also reside (Ref. 3).
Most functional agents belonging to level three reside in a
ground command and control center, or a manned vehicle in
the team. Some functions and services might also be
integrated onboard the UAVs with substantial computation
and communication capabilities. They function primarily as
decision support tools for the commander of the mission.
The middle level with local situation awareness is
responsible for the planning, re-planning and management of
missions (or sub-missions) allocated to a single UAV. A
level two mission plan is essentially a sequenced set of
orders for level one. A task library is used when a higher
level task is decomposed into lower level orders. The local
situation awareness agent monitors the world situation for
obstacles and threats. A synergy between the formation
control agent and its obstacle avoidance counterpart is
required in order to avoid obstacles (stationary or moving)
and collisions. At the formation level, the UAV swarm must
avoid vehicle collisions as well as moving or stationary
targets. The system configuration must assure the global
asymptotic stability of the coordinated formation control
strategy to be adopted so that vehicles meet safety
requirements without deviating substantially from their
planned mission/trajectory objectives. At the vehicle level,
guided by the obstacle avoidance agent, each UAV must be
capable of avoiding moving and stationary threats/obstacles
while performing its assigned task (Ref. 13). The Fault
Detection and Identification (FDI)/reconfigurable control
agent detects fault conditions and activates control
reconfigurations. The formation control agent is discussed in
detail in a later section of this paper.
Level one with behavioral knowledge is the level closest to
the physical dynamics and instrumentation of the UAV. It
encapsulates the vehicle’s physical control systems, weapon
system, communication mechanism, and sensing capabilities
for upper level agents, allowing it to deal with higher level
linguistic commands, such as “move”, “search”, “shoot”,
and “communicate”. It is responsible for executing these
tasks by translating them into physical commands, set points,
way points, etc. The agents belonging to this level are
designed to support heterogonous UAV models. They are
not necessarily knowledgeable of any plans, strategy, or
teamwork.
Differential Game Approach to Formation
Control
Differential Game Theory was initially used to determine
optimal military strategies in continuous time conflicts
governed by some given dynamics and constraints (Ref. 14).
One such application is the so-called Pursuit Game in which
a pursuer has to collide with an evading target. Naturally, in
order to solve such a problem it is advantageous to know the
dynamics and the positional information of both the evader
and the pursuer, that is, the Pursuit Game will be viewed as a
Perfect Information Game.
Formation Control as a Pursuit Game
The formation control problem can be regarded as a Pursuit
Game, except that, it is in general, much more complex in
terms of the combined dynamical equations, since the
system consists of n pursuers and n evaders instead of only
one of each. However, if the group of vehicles is viewed as
the purs uer and the group of desired points in the formation
as the evader, the problem is essentially reduced to the
standard but much more complex pursuit game.
Stability of the formation of vehicles is guaranteed if the
vehicles can reach their destination within some specified
time, assuming that the destination points are avoiding the
vehicles in an optimal fashion. It seems counterintuitive that
the destination points should be avoiding the vehicles
optimally, however if the vehicles can reach the points under
such conditions then they will always be able to reach their
destination.
As a consequence of our stability criterion, it is necessary
not only to determine the control strategies of the vehicles
but also the optimal avoidance strategies of the desired
points. Let us label the final control vector of the vehicles by
φ and the control final vector of the desired points byψ .
Then, the main equation which has to be satisfied is:
 r r 
φ ψ 
∑
min max  V j ⋅ f j ( x, φ ,ψ ) + G( x ,φ ,ψ )  = 0

(1)
 j 
which has to be true for both φ and ψ .
r
The f j ( x, φ,ψ ) term is the jt h dynamic equation governing
the system, and the V j is the corresponding Value of the
r
game. G( x, φ ,ψ ) is a predetermined function which, when
integrated, provides the payoff of the game. Notice, that the
only quantity that is not specified in the equation is the
V j term.
From the main equation it is possible to determine the
retrograde path equations (RPEs), which will have to be
solved to determine the actual paths traversed by the
vehicles in the formation. However, initial conditions of the
retrograde path equations will have to be considered in order
to integrate the RPEs. These initial condition requirements
provide us with the ability to introduce tolerance boundaries,
within which we say that the formation has settled. Such
boundaries naturally add complexity to the problem,
however they also provide a framework for positional
measurement errors.
The above formulation suggests a way for approaching the constant distance separating the two desired points, and that
solution to differential game. However, how does one ensure the formation can only perform translations and not rotations
that inter-vehicle collisions are avoided? To ensure this, it is in the three dimensional space. Hence the dynamic equations
necessary to consider the payoff function determined by the become:
r
integral of G( x, φ ,ψ ) . As an example, if we simply seek that x& d = vxd
the vehicles must reach their goal within a certain time τ, v& xd = Fd cos(ψ 1 ) sin(ψ 2 ) − kd ⋅ v xd
r
then G( x, φ ,ψ ) = 1. This can be verified by evaluating y& d = vyd
τ
r v& yd = Fd sin(ψ 1 ) sin(ψ 2 ) − kd ⋅ v yd
∫ G(x ,φ ,ψ )dt = τ . Hence, we have restricted our solutions
0 z& d = v zd
to the initial vehicle deployment, which will ensure that the v& zd = Fd cos(ψ 2 ) − kd ⋅ v zd
vehicles will reach the desired points in τ time. However, if In the above dynamical systems, the k i and kd factors are
r
G( x, φ,ψ ) is changed to penalize proximity of vehicles to simply linear drag terms to ensure that the velocities are
one-another, only initial conditions that ensure collision free bounded, and the Fd and Fi terms are the magnitudes of the
trajectories will be valid. applied forces. Figure 4 shows the coordinate system and the
associated angles.
r
However, G( x, φ ,ψ ) does not provide the means to perform
the actual collision avoidance, but merely limits the solution
space. So, in order to incorporate collision avoidance into
the controller, one can either change the value function or
add terms to the system of dynamic equations.

Limitations
Due to the solution’s dependence upon the dynamic
equations of the vehicles, some limitations of the differential
game approach naturally arise. As an example, consider the
case where it is desirable to replace one type of vehicle with
another in real time. Such systems can be accommodated
only if the new type of vehicle is capable of performing all Figure 4: Definition of Angles
the maneuvers assigned to its predecessor.
Substituting the dynamical equations into the main equation
(1), we obtain the following expressions:
min [ F1 ⋅ (Vvx1 ⋅ cos(φ1) ⋅ sin(φ2 ) +
Another issue that may arise, is the existence of a closed
form solution to the Retrograde Path Equations. For some φ
systems of dynamical equations it is simply not possible to
find the solution to the differential equations.
)
Vvy1 ⋅ sin(φ1) ⋅ sin(φ2 ) + Vvz 1 ⋅ cos(φ2 ) +
F2 ⋅ (Vvx 2 ⋅ cos(φ3 ) ⋅ sin(φ 4 ) +
Two-Vehicle Example Vvy 2 ⋅ sin(φ3) ⋅ sin(φ4 ) + Vvz 2 ⋅ cos(φ4 ) ] ) (2)
In order to illustrate some of the advantages and And
max [ Fd ⋅ (Vvxd ⋅ cos(ψ 1 ) ⋅ sin(ψ 2 ) +
disadvantages with the differential game approach to
formation control, consider the following system of simple ψ
point “Helicopters”, that is, points that can move in three
dimensions governed by the following dynamic equations:
Vvyd ⋅ sin(ψ 1) ⋅ sin(ψ 2 ) + Vvzd ⋅ cos(ψ 2 ) ] )
x& i = vxi To obtain the control law that results from the max-min
solution of equation (2), the following lemma is used:
v& xi = Fi cos(φ 2i −1) sin(φ 2i ) − ki ⋅ v xi Lemma 1:
y& i = vyi Let a, b ∈ ℜ :
v& yi = Fi sin(φ2i −1 ) sin(φ2i ) − ki ⋅ v yi ρ= a2 + b2
z&i = v zi Then
v& zi = Fi cos(φ2i ) − ki ⋅ vzi max( a ⋅ cos(θ ) + b ⋅ sin(θ ))
θ
Where i = 1,2. is obtained where
The two desired “points” are described by one set of
dynamic equations. This simply implies that there is a
 a b the analysis is performed on the actual position and velocity
cos(θ ) = , and sin(θ ) = differential equations.
ρ ρ
and the max is
ρ Furthermore, it should also be noted that this solution
closely resembles the isotropic rocket pursuit game
described in (Ref. 14). This is due to the fact that the
By combining Lemma 1 with Equation 2, the following
dynamic equations are decoupled, and hence working within
control strategy for vehicle 1 is found:
a three-dimensional framework will not change the problem
Vvx1 Vvy1
cos(φ1 ) = − , sin(φ1 ) = − considerably.
ρ1 ρ1
V ρ Simulation Results without Collision Avoidance
cos(φ2 ) = − vz1 , sin(φ2 ) = − 1
ρ2 ρ2
Where From the closed form expression of the control presented in
the previous section, it is obvious that the optimal strategies
ρ1 = Vvx1 + Vvy1
2 2
are in fact bang-bang controllers. Since the forces in the
and system are not dependent on the proximity of the vehicles to
the desired points, there will always exist some positional
ρ2 = Vvx21 + Vvy21 + Vvz21 error. It is however possible to resolve this problem simply
Similar results are obtained for vehicle 2. For the optimal by switching controllers at some error threshold, or
avoidance strategy of the desired points, we obtain the introducing terms that minimize the force terms F1 and F2 as
following: the vehicles approach the desired points.
V Vvyd
cos(ψ 1 ) = + vxd , sin(ψ 1 ) = +
ρd 1 ρ d1
Vvzd ρ
cos(ψ 2 ) = + , sin(ψ 2 ) = + d 1
ρd 2 ρd2
From this , we see that the retrograde equations have the
following form:
o V
v x1 = − F1 ⋅ vx1 + k1 ⋅ v x1
ρ2
o
x1 = −v x1
o
Vx1 = 0
o
V vx1 = V x1 − k1 ⋅ Vvx1 Figure 5: Two-Vehicle Simulation with Sufficient Vehicle
Velocities
For this example, the final value will be zero, and occurs
when the difference between the desired position and the
actual position is zero. Naturally, to obtain a more general
solution, a solution manifold should be used; however, in
order to display the utility of this approach, the previously
mentioned final conditions will suffice. The closed form
expression of the value function is then of the form:
1 − e − k1t
Vvx1 = ( x1 − xd ) ⋅
k1

It should be noted that the above analysis could be

performed on a reduced set of differential equations, where
each equation would express the differences in distance and
velocity, and hence reduce the number of differential
equations by a factor of 2. However, for the sake of clarity, Figure 6:Two-Vehicle Simulation with Insufficient Vehicle
Velocities
The above plot shows the tracking capabilities of the derived
controller. The two vehicles are attempting to follow two
parameterized circular trajectories with a radius of three. In
Figure 5 the vehicles can move quickly enough to actually
reach the desired trajectories, while in Figure 6 the velocities
of the vehicles are not sufficient to reach the desired
trajectories. In the latter case, the vehicles simply move in a
smaller circle, which ensures that the error remains constant.
Let us consider the magnitude of the tracking errors for these
two cases:
Figure 7: Position Error with Sufficient Vehicle Velocities
Figure 8: Position Error with Insufficient Vehicle Velocities
Notice, that the error curves in Figure 7 are fairly smooth up
until the desired trajectory is reached. However, as soon as
the desired trajectory is reached, the magnitude of the error
changes quite rapidly. This is due to the bang-bang control
law discussed earlier. When the system attempts to correct
for a small positional deviation, the force vector applied may
cause an overshoot, and hence the magnitude of the velocity
vector becomes somewhat irregular. This phenomenon is
naturally not present in the second case, since the vehicles
never reach the desired trajectories and consequently the
trajectories in Figure 8 remain smooth.
Collision Avoidance
Collision avoidance was added to this problem, simply by
observing that the control is only dependent on the value
function V. Hence, if we simply design the value function
such that it ensures that the two vehicles moves away from
one another when they come too close to each other, the
controller will in essence avoid collisions. However, since
the collision avoidance is not accomplished by changing the
dynamic equations that we base the value function on, the
convergence time is no longer guaranteed.
Figure 9: Two-Vehicle Simulation with Conflicting Goals
The simulation below displays a scenario where the
formation is obviously unstable. The two vehicles are
attempting to reach the same trajectory, however due to the
addition of collision avoidance the desired formation will
never be reached.
Figure 10: Position Error with Conflicting Goals
Notice, that the 3D position plot in Figure 9 seems to
suggest that the two vehicles actually reach the desired
trajectory. However, the position error plot shown in Figure
10 verifies that the desired safety region of two units is never
compromised, that is, the second vehicle is simply trailing
the first along the desired trajectory.
Simulation Results with Additional Vehicles
Since the dynamic equations used to derive the controller
have no inter-vehicle dependence, it stands to reason that the
system is scaleable to n-vehicles. Figure 11 shows a
simulation with ten vehicles following a circular trajectory
with a radius of six units. The circles overlap slightly,
however this does not cause conflicts to arise since no two
vehicles will be in those regions at the same time. However,
the overlap causes the system to converge slower than the
two-vehicle system, as shown in Figure 12. This decrease in
convergence rate is due to conflict resolution on approach to
the desired trajectories. Such resolution is clearly displayed
around time index 150, where one of the vehicles has to
leave the desired path to ensure that collision is avoided, and
consequently increases its corresponding positional error. It
should be noted that, after time index 300, the system has
settled on the desired trajectory.
Figure 11: Ten-Vehicle Simulation without Conflicting
Goals
Figure 12: Convergence of Ten-Vehicle System
Figure 13: Plot of Number of Collisions as Vehicles Attempt
to Reach the Same Trajectory
The approach seems to perform remarkably well under such
well-defined desired trajectories, however what would
happen if they were all supposed to reach the same trajectory?
Under such impossible conditions the proposed system will
not guarantee collision avoidance.
As shown in Figure 13, the number of collisions appears
initially to be increasing drastically with the number of
vehicles. However, when the number of vehicles is large,
the addition of extra vehicles seems not to impact the
number of collisions significantly.
Based on the simulation, it is clear that well-defined
trajectories are not just vital to the stability of the formation
but also essential to the safety of the vehicles. In the
simulation of the system with increasing number of vehicles,
a collision was defined as a point in time when the distance
between two vehicles was less than or equal to one.
Conclusion
• A hierarchical multi-agent architecture for the
coordinated control of multiple UAVs has been
presented in this paper.
• The formation control problem is viewed as a Pursuit
Game of n pursuers and n evaders. Collision avoidance
is achieved by designing the value function so that it
ensures that the two vehicles move away from one
another when they come too close.
• By viewing the formation control problem as a
differential game, important performance information
about the formation can be determined, for example, the
existence of solutions for any given set of initial
conditions, the time to reach the target and whether a
designated formation flight path is reachable. However,
the mathematical analysis required to obtain such
 information is considerable, and in some cases obtaining
a closed form expression for the control law may be
impossible. Moreover, the analysis of one formation of
vehicles cannot always be translated onto another
formation with different dynamics. The lack of a closed
form solution could be remedied by using numerical
methods; however, the dependency on the individual
vehicle dynamics seem to be the price that has to be
paid to obtain the performance measures mentioned
above.
References:
1. Office of the Secretary of Defense (Acquisition,
Technology, & Logistics), Air Warfare. "OSD UAV
Roadmap 2002-2027." December 2002.
2. Sousa, J. B., and Pereira, F., "A framework for
networked motion control," Proceedings of the 42nd
IEEE Conference on Decision and Control, pp. 1526-
1531, Hawaii, USA, December 2003.
3. Howard, M., Hoff, B., Lee, C., "Hierarchical Command
and Control for Multi-agent Teamwork", Proceedings of
5th Intl. Conf. on Practical Application of Intelligent
Agents and Mult-Agent Technology (PAAM2000), pp.
1-13, Manchester, UK. Apr. 10, 2000.
4. Baras, J. S., Tan, X., and Hovareshti, P., "Decentralized
Control of Autonomous Vehicles," Proceedings of the
42nd IEEE Conference on Decision and Control, pp.
1532-1537, Hawaii, USA, December 2003.
5. Dunbar, W. B., Murray, R. M., "Model Predictive
Control of Coordinated Multi-Vehicle Formation,"
Proceedings of the 41st IEEE Conference on Decision
and Control, pp. 4631-4636, Las Vegas, USA,
December 2002.
6. Voos, H., “Market-based Algorithms for Optimal
Decentralized Control of Complex Dynamic Systems ,”
Proc. of the 38th IEEE Conference on Decision and
Control, Vol.40, pp.3295-3296, Phoenix, AZ. 1999.
7. Clearwater, S. H. E., "Market-Based Control: A
Paradigm for Distributed Resource Allocation,"
Singapore: World Scientic, 1996.
8. Walsh, W., and Wellman, M., “A Market Protocol for
Decentralized Task A llocation,” Proc. of the 3rd
International Conference on Multiagent Systems , 1998.
9. Engelbrecht, W. R., Shubik, M. , and Stark, R. M.,
"Auctions, Bidding, and Contracting: Uses and Theory,"
New York, NY: New York University Press, 1983.
10. Bertsekas, D., “Auction Algorithms for Network Flow
Problems: A Tutorial Introduction,” Com-putational
Optimization and Applications, Vol.1, pp.7-66, 1992.
11. Vachtsevanos, G., Kim, W., Al-Hasan, S., Rufus, F.,
Simon, M., Schrage, D., and Prasad, J. V. R., "Mission
Planning and Flight Control: Meeting the Challenge
with Intelligent Techniques," Journal of Advanced
Computational Intelligence, Vol. 1, (1), pp. 62-70, Oct.,
1997.
12. Al-Hasan, S. and Vachtsevanos, G., "Intelligent Route
Planning for Fast Autonomous Vehicles Operating in a
Large Natural Terrain," Journal of Robotics and
Autonomous Systems, Vol. 40, pp. 1-24, 2002.
13. Al-Hasan, S., Vachtsevanos, G., "A Neural Fuzzy
Controller for Moving Obstacle Avoidance," Third
International NAISO Symposium on Engineering Of
Intelligent Systems , Malaga, Spain, September 24 - 27,
2002.
14. Isaacs, R., "Differential Games: a Mathematical Theory
with Applications to Warfare and Pursuit, Control and
Optimization." New York: John Wiley and Sons, Inc.,
1965.