See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/241885390

12 Artificial Intelligence for Space Applications

Article · January 2007

DOI: 10.1007/978-1-84628-943-9_12

CITATIONS READS

13 2,453

2 authors, including:

Dario Izzo
European Space Agency
208 PUBLICATIONS   1,816 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Machine learning for interplanetary trajectories View project

Differential Intelligence View project

All content following this page was uploaded by Dario Izzo on 10 October 2014.

The user has requested enhancement of the downloaded file.

12
Artificial Intelligence for Space Applications

Daniela Girimonte and Dario Izzo

European Space Agency, Advanced Concepts Team, ESTEC, EUI-ACT,

Keplerlaan 1, 2201 AZ Noordwijk, The Netherlands,
daniela.girimonte@esa.int, dario.izzo@esa.int

Summary. The ambitious short-term and long-term goals set down by the various
national space agencies call for radical advances in several of the main space engi-
neering areas, the design of intelligent space agents certainly being one of them. In
recent years, this has led to an increasing interest in artificial intelligence by the
entire aerospace community. However, in the current state of the art, several open
issues and showstoppers can be identified. In this chapter, we review applications
of artificial intelligence in the field of space engineering and space technology and
identify open research questions and challenges. In particular, the following topics
are identified and discussed: distributed artificial intelligence, enhanced situation
self-awareness, and decision support for spacecraft system design.

12.1 Introduction
In the second half of 2003, the European Space Agency (ESA) delivered
a roadmap, in the framework of the Aurora program, to bring humans to
explore Mars within the next few decades [MO03]. The plan included the
successful implementation of several flagstone missions as stepping stones
for achieving this final ambitious goal. A few months later, with the vision
delivered by U.S. president George W. Bush, the National Aeronautics and
Space Administration (NASA) also started to draft plans for the manned
exploration of Mars [Bus04]. Their vision included the establishment of a
human base on the Moon, among several other advanced preparatory steps.
The return of humans to the Moon and a future manned mission to Mars
therefore seem to be likely achievements we may witness in the next few
decades. At the same time, even more ambitious plans and missions are being
conceived by farsighted researchers who dream about the exploration and
colonization of even farther planets.
In the framework of these more or less concrete future scenarios, the
consolidation of artificial intelligence methods in space engineering is certainly
an enabling factor. As an example, the reader may think of a future mission to
Mars. This will probably be constituted by a large number of heterogeneous
236 Daniela Girimonte and Dario Izzo

space agents (intended to be satellites, humans, robots, modules, sensors,

and so on). In such a scenario, the round-trip communication delay time,
depending on the relative positions of Mars and Earth, would range from 6.5
minutes to 44 minutes approximately. Besides, communication with Earth
would not be possible at all during a 14 day period every Mars synodic
period (approximately 2.1 years). Clearly, for such a mission to happen, the
single space agents must be able to make autonomous decisions, to interact
harmoniously with each other, and to be able to determine their own health
status so as to properly plan their actions. Unfortunately, if we take a look
at the current state of the art of artificial intelligence applications in space
engineering, we can identify several open issues and showstoppers. Actually,
it seems that we are far away from the desirable situation in which these
methods can be considered as off-the-shelf tools available to space engineers.
This chapter is addressed to the artificial intelligence community in order
to create an awareness of the many open research questions and challenges
in the space engineering community. In order to achieve this task, the
chapter focuses on a few niche applications only, namely distributed artificial
intelligence for swarm autonomy and distributed computing, and enhanced
situation self-awareness and decision support for spacecraft system design.
Our survey aims to give the reader a general overview of these topics by
pointing out some of the relevant activities within the international space
community and as such is not intended to cover the entire array of all artificial
intelligence applications in space. For example, we deliberately omitted in
this discussion research on automated planning and scheduling, which is
traditionally the most studied field within artificial intelligence for space, and
we refer interested readers to other resources such as the proceedings of the
International Workshop on Planning and Scheduling for Space (e.g., 1997,
2000, 2004, and 2006).

12.2 Distributed Artificial Intelligence

At the end of the 1980s, the artificial intelligence community started wondering
whether intelligence had to be strictly related to reasoning. Failures in
constructing agents able to interact with the environment in real-time following
high-level decisions derived via symbolic reasoning led to a new approach
in the design of robot control systems: “behavior based” robotics [Bro91].
Starting from the simple observation that most of what we do in our daily
lives is not related to detailed planning but rather to instinctual reactions
to an unstable and changing environment, behavioral robotics introduced,
for the first time, the notion of “emerging” intelligence. Researchers were
forced to observe that, in some systems, intelligence could emerge from the
interaction with the environment and from indirect relations between system
parts and that, in general, intelligence could not always be easily located in
one particular part of the system studied. The idea of intelligence residing in a
 12 Artificial Intelligence for Space Applications 237

distributed form throughout an agent started the study of intelligent systems

made by more than one agent. Hence, “distributed artificial intelligence”
(DAI) developed as a discipline studying systems made up of a number of
diverse agents that despite their individuality are able to achieve common
global tasks. In the following sections, we mainly touch upon two topics of DAI
systems for space applications: swarm intelligence and distributed computing.

12.2.1 Swarm Intelligence

There is no common agreement on the definition of swarm intelligence.

Definitely a subcategory of distributed artificial intelligence, we define swarm
intelligence as the emerging property of systems made by multiple identical
and noncognitive agents characterized by limited sensing capabilities. This
definition, nearly a description of biological swarm intelligence, stresses the
necessity of having agents that interact locally with the environment and
between themselves. It may be argued that algorithms historically considered
at the center of swarm intelligence research, such as “particle swarm optimi-
zation” (PSO) [KE95], sometimes lack this property, which therefore should
not be required in the definition.1 Others would instead take the opposite
direction in requiring that the local interaction happen only indirectly through
an intentional modification of the environment (stigmergy [TB99]). Other
issues arise when trying to decide whether deterministic systems should be
excluded from the definition of swarm intelligence. These pages may not be a
suitable place to settle or discuss these issues, and we therefore ask the reader
to be forgiving should our view not be fully satisfactory to him or her.
Whatever definition one wishes to adopt, a number of features of swarm
intelligence are certainly attractive to the space engineering community. The
space environment typically puts stringent constraints on the capabilities of
single satellites, robots, or anything that needs to survive in space (space
agents). Space agents are particularly limited in terms of mobility (propellant-
and power-limited), communication (power-limited), and size (mass-limited).
At the same time, a high level of adaptability, robustness, and autonomy
is required to increase the chances of success of operating in a largely
unknown environment. Similar characteristics are found in the individual
components of a biological swarm. Moreover, a number of space applications
are naturally based on the presence of multiple space agents. The first
commercial application proposed and realized for satellite systems was that
of Arthur C. Clarke and was a satellite constellation providing global commu-
nication services by means of three satellites put in a geostationary orbit
[Cla45]. Since then, a large number of constellations have been deployed to
provide global communication, navigation, and Earth observation services.

1
The so-called social component in the PSO algorithm requires at each step for
each agent to know the best solution found so far by the entire swarm. Interagent
communication is, in this case, direct and unlimited in range.
238 Daniela Girimonte and Dario Izzo

More recently, the idea of a number of satellites flying in formation has been
used in a number of missions for applications ranging from x-ray astronomy
(XEUS) to differential measurements of the geomagnetic field (CLUSTER
II), space interferometry, the search for exoplanets (DARWIN), and others.
All these missions 2 are able to meet their requirements without making use of
an emerging property that can be regarded as swarm intelligence. On the other
hand, if available, swarm intelligence methods would represent an attractive
design option allowing, for example, achievement of autonomous operation
of formations. Simple agents interacting locally could be considered as a
resource rather than overhead. At the same time, one would be able to engineer
systems that are robust, autonomous, adaptable, distributed, and inherently
redundant. Besides, swarms allow for mass production of single components,
thus promising mission cost reduction, and represent highly stowable systems,
thus allowing reduced launch costs. Recently, these motivations led a number
of researchers to simulate some degree of swarm intelligence in a number of
space systems and to investigate their behavior.
Kassabalidis et al. [KEM+ 01] studied the routing problem in wireless
communication networks between satellites or planetary sensors. He applied
ant-inspired algorithms to achieve a great efficiency in networks that are
spatially distributed and changing over time. This type of research is targeted
at applications such as those being developed by the NASA sensorweb project
[CCD+ 05]. Distributed cooperative planning between satellites belonging to
the same constellation has also been studied, introducing swarm intelligence
at the level of coordinated planning [DVC05] (for a typical case study, see
Fuego, studied by Escorial et al. [ETR+ 03]). Recent work on intersatellite
communication in constellations observed the birth of emerging properties
from a more or less complex system of rules and behaviors [BT07] programmed
in the autonomous planners onboard the satellites. More generally, any
problem of autonomy for satellite constellations is a problem of distributed
artificial intelligence, where the possibility of communication between agents
(ISL-intersatellite links) or between an agent and a central planner (ground
station) is limited by the complex dynamics of the system and by the agent
design.
Another field where swarm intelligence provides a possibility to improve
current technology is that of relative satellite motion control. When a system
of many satellites has to move in a coordinated way, the control action
selected by each satellite may take into account the decisions made by the
others at different levels. The information exchanged with the other swarm
components is useful but not necessary to define the geometric and kinematical
representation of the time-varying environment, which will then influence
the satellite action selection. Many studies dealing with terrestrial robot
navigation [Kha86], with spacecraft proximity and rendezvous operations
[McI95], and self-assembly structures in space [McQ97] have taken the ap-

2
At the time of writing, CLUSTER II is the only one operational.
 12 Artificial Intelligence for Space Applications 239

proach of defining an artificial potential field to model the environment.

The control action is then chosen to follow the steepest descent of the
defined potential. Another approach to the action selection problem, based on
dynamic systems theory, was introduced by Schoner [SD92]. In this approach,
the state-space contains behavioral variables such as heading directions or
velocities. All the contributions given by each behavior are combined by
means of weighting parameters into a final dynamical system that defines
the course of behaviors that each agent will follow. The weighting parameters
can be evaluated by solving a competitive dynamic system operating on a
faster timescale. Recently, other approaches have been proposed, in particular
for space applications, attempting to obtain some degree of decentralized
coordination in a group of satellites. Lawton and Beard [RB04, LYB03]
introduced what they call a “virtual structure” method to design a decen-
tralized formation control scheme. Their method aims at reaching a unique
final configuration in which each satellite has its position preassigned. When a
swarm of homogeneous agents is considered and the task is given to acquire a
certain final geometry, the final positions occupied by each agent in the target
configurations should be chosen in an autonomous way and should be part of
the global behavior emerging from the individual tasks assigned. This result
is actually possible using a technique [IP07] developed at ESA and inspired
by swarm aggregation results [Gaz05] for terrestrial robots. Introducing a
behavioral component that accounts for the differential gravity typical of
orbital environments, the algorithm allows one to obtain, in a given countable
number of final formations, a swarm whose emerging behavior is the solution
of the target allocation problem and the acquisition and maintenance of the
final formation.
Figure 12.1 illustrates two examples of orbital swarms controlled by this
algorithm with the addition of a limited amount of hierarchy in the swarm to
allow the lattice formation [IP07]. Note that, in the first case, the swarm
autonomously selects its final arrangement from 2.81 · 1041 different final
possible configurations. In the second case, this number is 4.16 · 1029 . A
behavior-based control approach for satellite swarms has also been shown
to be useful in controlling highly nonlinear systems such as those derived by
introducing electrostatic interactions between swarm agents [PIT06].

12.2.2 Distributed Computing

A second example of distributed artificial intelligence with specific applications

to space systems, and in particular to trajectory design [IM05], is that
of distributed computing. The possibility of sharing the memory and the
computing resources of a large network of simple computers is clearly appealing
for any kind of application. On the other hand, not every problem is suitable
for being solved in a distributed computing environment. The problem struc-
ture has to be such as to allow its subdivision into packages that have little or
no dependency between each other. This requirement is the main limitation to
240 Daniela Girimonte and Dario Izzo

Fig. 12.1. Two examples of orbital swarms assembling a given structure (source:
[IP07]).

the use of distributed computing. The forthcoming sections introduce, briefly,

two examples of space applications suitable for distributed computations.

Analysis of Large Quantities of Data

The main purpose of most of the commercial satellites currently orbiting

Earth is to provide data. Satellites continuously download data to ground
stations in a nonprocessed format (usually, few data manipulations are made
by the not too powerful computers onboard satellites). ESA’s ENVISAT
satellite alone generates 400 terabytes of data each year [FGL+ 03]. The
data are then processed sequentially by computers and the results stored
again in mass memories together with raw ones. Over the years, these data
accumulate to the point that deletion is sometimes necessary (also due to
changes in storage technology). Sophisticated analysis of these datasets can
take as long as years to complete, often making the analysis itself obsolete
before it has even been concluded. Distributed computing therefore becomes
a useful tool to allow efficient use of satellite data, the main asset of the
space business. Earth observation data coming from European satellites have
already been made available in a computer grid [FGL+ 03], sharing processing
power, memory storage, and processed data. A dedicated generic distributed
computing environment that uses the idle CPU time of ESA internal personal
computers has also been tested already [IM05] on problems such as ionospheric
data processing and Monte Carlo simulations of constellation architectures
 12 Artificial Intelligence for Space Applications 241

[IMN05]. In these types of applications, as no dependency is present between

the different parts of the computations, little distributed artificial intelligence
is actually present. The huge problem is just divided into small isolated
subproblems that, in turn, are solved by different machines located in various
parts of a common network. From the technological point of view, the
challenges in these types of distributed computations (and the part where
artificial intelligence could play a role) are mainly in the coordination of
network traffic, in resource sharing, and in the reconstruction of the whole
solution from the different parts returned by the different machines.

Distributed Solving of Global Optimization Problems

Distributing global optimization tasks over a large network of computers is

certainly more elaborate, as it introduces a dependency between the different
computations. Global optimization problems can be found everywhere in
industrial processes. Many of the issues engineers face during spacecraft design
are global optimization problems. Most notably, global optimization seems to
be essential in preliminary trajectory optimization [MBNB04]. Essentially,
this can be considered in the rather generic form

min : f (x)
subject to : g(x) ≤ 0

with x ∈ U ⊂ Rn . The problem dimension n depends on the type of trajectory

considered and can be as low as 2 but also on the order of thousands.
In order to visualize, for the reader, the problem of trajectory optimization,
Figure 12.2 illustrates an example of an optimized interplanetary trajectory.
Since the first applications of evolutionary strategies to trajectory design
[RC96], heuristic optimization techniques such as differential evolution, simu-
lated annealing, particle swarm optimization, and genetic algorithms have
proven to be quite effective in providing preliminary solutions to trajectory
problems [BMN+ 05, DRIV05]. The complete automatization of the optimi-
zation process, however, is not yet possible, as the existing algorithms are
incapable of replacing the acute reasoning necessary to locate the best
possible transfer between celestial bodies. A recent attempt to capture some
expert knowledge and to use it to prune the search space in a trajectory
problem, called “multiple gravity assist” (MGA), has managed to reduce
the MGA problem complexity to polynomial time [IBM+ 07]. In other cases,
NP-complexity cannot be avoided, and the global optimization of an inter-
planetary trajectory may be untractable for a single machine. Fortunately,
global optimization algorithms such as evolutionary algorithms and branch-
and bound-based techniques are suitable for distributed environments [GP02,
AF98], drastically improving the performance of the search and thus allevi-
ating the “curse of dimensionality”. A first attempt to use this possibility
in a spacecraft trajectory optimization problem has been performed by ESA
242 Daniela Girimonte and Dario Izzo

9
x 10

Saturn Encounter

1
Impact Point

0.5

0
km

−0.5

−1

Jupiter encounter
−1.5
−1.5 −1 −0.5 0 0.5 1 1.5
km 9
x 10

Fig. 12.2. An optimized Earth-Venus-Earth-Venus-Earth-Jupiter-Saturn-asteroid

trajectory.

researchers, who solved a complicated MGA transfer distributing a differential

evolution algorithm in a small number of personal computers. The problem
solved3 using the distributed version of differential evolution was inspired by
the 1st Global Trajectory Optimization Competition (GTOC1), an annual
event established in 2005 to make international research groups compete to
find the best solution to the same trajectory design problem. Depending
on the type of spacecraft one is considering (the main difference being the
possibility of having impulsive or continuous velocity changes), the mission
goal (destination orbit and celestial body), and the launch window considered,
the trajectory optimization problem’s dimension and complexity vary a lot.
As in many other fields, for trajectory optimization, too, there is no available
algorithm that outperforms all others. Consequently, this often leaves one to
try different techniques until finally the algorithm that in a particular problem
performs best is found. In the attempt to make the entire problem-solving
process entirely distributed, a novel approach is that of Vinko et al. [VIP07],
who consider a central server and a number of clients, which evolve demes
(subpopulations) extracted from a larger population stored in the server,
3
There is actually no mathematical guarantee that the solution found is the global
optimum, but the experiment improved previous solutions by approximately 10%.
 12 Artificial Intelligence for Space Applications 243

which then takes care of the reinsertion of the demes. According to the results
returned by the various clients in each given phase of the optimization process,
the server updates the probabilities to allocate a given algorithm for the next
subsearch request to a client. The resulting global optimization environment
is able to understand and select the best-performing algorithm in each phase
of the solution of a problem. A preliminary version of this intelligent server
is being developed and tested [VIP07] for the final purpose of being able to
automatically carry out the whole trajectory optimization process without
any expert supervision.

12.3 Enhanced Situation Self-Awareness

Ideally, a spacecraft should be able to perform autonomous actions, determine

its own health status, and eventually make decisions based on this enhanced
self-awareness. Unfortunately, real space missions are instead strongly depen-
dent on the ground segment and on the flight engineers who monitor the
enormous amount of telemetry data sent back to Earth during spacecraft
operation. This procedure, which requires large numbers of human experts,
is of course cumbersome and time-consuming. Sometimes, it might take days
before the data are processed, decisions are made, and uploaded commands
reach the spacecraft, whereas during critical mission phases such as the
launch, information must be processed and decisions make within seconds.
Furthermore, humans are not always able to recognize anomalous situations,
especially when these involve complex relationships among large numbers
of variables. Autonomous systems for enhanced situation self-awareness are
therefore a very important research topic in spacecraft engineering.
Classically, two major approaches can be described: model-based methods
and data-driven (model-free) methods. Model-based methods use models of
the hardware and the physical processes to track the states of the system
and detect deviations from nominal behavior. These models are sometimes
very expensive to produce because they largely depend on expert knowledge.
Moreover, when applied to very complex systems such as spacecraft, they
might fail to reproduce all the possible off-nominal modes for which accurate
models are lacking most of the time.
On the other hand, data-driven approaches, based on data mining and
machine-learning techniques, are not based on a physical system but rather
on models that are inferred from the telemetry data (e.g., temperature sensor
data). Many activities in this field are being carried out in the framework
of the Integrated Vehicle Health Management (IVHM) program of NASA
Ames Research Center for the Second Generation Reusable Launch Vehicle
(RLV), crew, and cargo transfer vehicles [IVH06]. This program is dedicated
to the development of highly integrated systems that will include advanced
smart sensors, diagnostic and prognostic software for sensors and components,
model-based reasoning systems for subsystem- and system-level managers,
244 Daniela Girimonte and Dario Izzo

advanced onboard and ground-based mission and maintenance planners, and

a host of other software and hardware technologies. These hardware and
software technologies will provide both real-time and life-cycle vehicle health
information, which will enable decision making.

12.3.1 Data-Driven Approaches

The application of data-driven approaches to flight time-series analysis is

being researched extensively by the space engineering community for the
autonomous identification of suspicious trends that might lead to malfunctions
or losses. Only the preventive detection of these trends might allow the
ground systems or the intelligent planner and scheduler of the spacecraft
to take corrective actions. Most of the data-driven approaches used in daily
spacecraft operations are based on unsupervised learning techniques since in
safety-critical applications, such as space engineering, it is usually impossible
to collect exhaustive datasets for the representation of all possible fault
modes. Therefore, most of these methods and algorithms can detect anomalies
and off-nominal trends but leave to the flight control operator the delicate
task of interpretation. The forthcoming paragraphs introduce a few of these
approaches. Far from being an exhaustive list, we intend to give the reader a
flavor of some work done in this field.
In [Ive04], the authors propose an “inductive monitoring system” (ISM) to
detect off-nominal behaviors. Flight data of past missions are used as training
data for a clustering algorithm (i.e., K-means and density-based clustering)
that identifies nominal behavior areas (the clusters) in the n-dimensional
data space, where n is the number of sensor readings. The clusters, which,
according to the authors, represent the ISM knowledge base, can be used
for the real-time detection of anomalous behavior during a new flight. Once a
new measurement vector is received, the knowledge base returns the cluster to
which the vector would belong (according to some cluster limit, preidentified
after training). When the membership in a specific cluster cannot be detected,
the distance to the closest cluster (with respect to Euclidean metrics in the
n-dimensional space) will give the control operator an idea of the system’s
deviation from its nominal behavior as represented by the training data. The
algorithm is tested successfully on the data collected during mission STS-107
of the Columbia space shuttle, which exploded during reentry because of a
breach in its thermal protection system [Geh03]. An approach very similar to
the one just introduced is presented in [Sch05]. In this work, an unsupervised
detection algorithm named Orca, developed by the authors on the basis of
the nearest-neighbor approach, is applied to the test data of the space shuttle
main engine and of a rocket engine stand.
The K-means clustering algorithm is also used in [VLFD05] on the space
of the features extracted from the time series collected from past missions.
The authors here make an attempt to find specific relations between fault
occurrence and the trend of the parameters by inferring association rules
 12 Artificial Intelligence for Space Applications 245

from data by means of the a priori algorithm. Therefore, the fault occurrence
data must be part of the training set so that the algorithm can be trained to
recognize future similar events. Unexpected fault modes therefore cannot be
detected by the algorithm.

12.3.2 Model-Based Approaches

Most of the model-based approaches for enhanced situation awareness that

have been researched and developed in space systems in recent years used
as a reference the Livingstone model-based diagnosis engine [WN96] and its
successors Livingstone 2 (L2) and Livingstone 3 (L3). Livingstone flew on
Deep Space 1, and L2 has been uploaded to the Earth Observing One (EO-1)
satellite [HSC+ 04, CST+ 04] for the “autonomous sciencecraft experiment”
(ASE), which provides an onboard planning capability. The task of these
diagnosis engines is to predict nominal state transitions initiated by control
commands monitoring the spacecraft sensors and, in the case of failure, isolate
the fault based on the discrepant observations. Fault detection and isolation is
done by determining a set of component modes, including most likely failures,
which satisfy the current observations.
L3 is the most recent and advanced of these architectures and consists
of three main components. The “system model” stores the model of the
system and is responsible for tracking the modes of operation of the different
components and determining the constraints that are valid at any point in
time. The “constraint system” serves the role of tracking the overall system
behavior using constraint programming techniques. It receives constraints
from the system model indicative of the current configuration of the system
and propagates these constraints to try to assign consistent values to variables
in the system. When observations are different from propagated values for
corresponding components, the “candidate manager” is responsible for gener-
ating candidate faults that resolve all the conflicts and that can possibly
explain all of the inconsistencies. In order to deal with uncertainties, the
dynamic behavior of the system is tracked through Bayesian approaches such
as “particle filtering” in order to assign posterior probability distributions to
the candidate faults [NDB04, NBB04].
Bayesian approaches are also used in [GIB06], where the authors present
the preliminary results of dual filtering techniques for the detection of possible
variations of the thermal properties of the spacecraft that result from vari-
ations of its physical properties and for determining a complete thermal
mapping of the system. System and sensor uncertainties are taken into
account in the lumped parameter modeling of the thermal system, and a
dual unscented Kalman filter is run on the stochastic model in an alternating
optimization fashion to estimate the thermal state and coefficients of the
resulting thermal network from the readings of a few strategically placed
thermal sensors. Events such as faults can be detected by the dual filter as
well as new values of system parameters (e.g., radiative couplings) that result
246 Daniela Girimonte and Dario Izzo

from a variation of the spacecraft geometry (e.g., from the deployment of

antennas, solar panels, etc.). This method would be particularly attractive in
networks whose state and parameters can be estimated by the filter using a
minimal amount of readings. The relation between the network topology and
this minimal number is therefore an issue strictly related to the observability
of the system, which is here approached using graph theory.

12.4 Decision Support for Spacecraft System Design

As the complexity of space systems increases, innovative approaches to system
design are needed to allow assessment of the largest possible number of
design concepts at an early stage. In space system design, several disciplines
corresponding to all different subsystems4 must be considered, and the overall
spacecraft is the result of a “multidisciplinary design optimization” (MDO)
[BS02, Roy96]. MDO can be described as a methodology for the design of
systems where the interaction between several disciplines must be considered
and where the designer is free to significantly affect the system performance
in more than one discipline. In this sense, the space design process is an
integrated optimization 5 that receives as inputs the mission requirements in
the form of constraints and produces as output an optimal design.
In the classical approach to MDO, each specialist would prepare a sub-
system design relatively independently from the others using stand-alone
tools. Design iterations among the different discipline experts would take
place in meetings at intervals of a few weeks. This well-established approach
has the drawbacks of reducing the opportunity to find interdisciplinary
solutions and to create system awareness in the specialists. A considerable
step toward a multidisciplinary approach in the early phases of space system
design has been achieved through an MDO based on concurrent engineering,
where a sequential iterative approach to system design is replaced by a
parallel and cooperative approach. Design facilities where these methodologies
are implemented are, among others, the ESA Concurrent Design Facility
[BMO99], the NASA Goddard Integrated Mission Design Center [KMSR03],
and the Concept Design Center at The AeroSpace Corporation [ADL98].
In these concurrent MDO approaches, however, the subsystem experts
are the core of the decision process of the design. Over the last couple
of years, much research has been dedicated to the achievement of decision
support systems or that of autonomous system design methods, which try to
capture the reasoning of the experts toward an optimal and robust design.
4
A spacecraft is constituted by the following subsystems: attitude determination
and control, telemetry tracking and command, command and data handling,
power, thermal structures and mechanisms, and guidance and navigation [WL99].
5
The term optimization is not used here in the strict mathematical sense but rather
to indicate any procedure that aims to find a solution that is either optimal or
suboptimal.
 12 Artificial Intelligence for Space Applications 247

Therefore, the spacecraft design started to be viewed as the solution of an

optimization problem under constraints: given a set of decision variables D
(e.g., the dimension of solar arrays) and a set of constraints C on D (e.g., their
volume and mass), the constrained optimization algorithm looks for the values
of D that minimize or maximize an objective function F (X) subject to C.6
However, finding the optimal design point was revealed to be a very difficult
task, and traditional global optimization approaches most of the time fail to
find the global optimum in the design space [FCM+ 97]. To tackle this problem
in spacecraft design, a quite common approach is based on the employment of
heuristic solvers. The Jet Propulsion Laboratory implemented an optimization
assistant (OASIS) that depending on the design problem selects and tunes
either a genetic algorithm or a simulated annealing algorithm [FCM+ 97].
The goal of OASIS was to facilitate rapid “what-if” analysis of spacecraft
design by developing a spacecraft design optimization system that maximizes
the automation of the optimization process and minimizes the amount of
customization required by the user. More recently, evolutionary algorithms
have been used to evolve the design of the antenna that flew on NASA’s Space
Technology 5 (ST5) mission [HGLL06] and for trajectory design as discussed
in the previous section.
The problem of tackling the conflicting situations that might emerge
during the system design activity when interests from different disciplines
must be harmonized in the same project or when different goals must
be reached within the same mission has been studied in [AFA+ 04]. The
neighborhood approach aims at finding by means of dedicated heuristics a
set of “paretian” solutions at the system level. To efficiently reduce the total
number of such solutions to a small subset that is to be considered “optimal”
from the point of view of conflict reduction, “game theory” and “multicriteria
decision analysis” are used.
Other approaches to autonomous space system design look not only at the
achievement of an optimal design but also at its robustness with respect to
uncertainties of the design variables and models involved in the design.7 In this
framework, the most common approach in space system design is essentially
based on safety margins and expert knowledge. The safety margins, which
are the most conservative way of handling uncertainties, identify the worst
possible conditions that might be encountered during the operational phase
in order for the resulting design to be adequate. Probabilistic approaches have
been introduced in space system design as a consequence of the Challenger
accident in 1986 [Fey86] and are essentially based on “probabilistic risk
analysis” [PF93]. However, in general, the probability of infeasibility for a
given design cannot be determined reasonably without knowing the joint

6
In the case of spacecraft design, the objective function is most of the time the
cost, which is ultimately proportionally linked to be the spacecraft’s total mass.
7
For an extensive qualitative and quantitative overview of these uncertainties, the
reader may consult [Thu05].
248 Daniela Girimonte and Dario Izzo

distribution of the uncertain variables or having sufficient amounts of data

samples from past observations. Sometimes, the probability model assumptions
can be replaced by deterministic data, for which a rigorous worst-case analysis
could be performed by using numerically reliable tools, such as verified
interval calculations. In the most recent literature on system design under
uncertainties, design variables are modeled by a range of values (intervals),
by membership-degree functions of fuzzy sets [LF02], or by evidence theory
[CCV07]. The European Space Agency’s Advanced Concepts Team is assessing
a promising new approach for an autonomous and robust design based on
the concept of clouds [Neu04, DP05]. Clouds capture useful properties of
the probabilistic and fuzzy uncertainties, enabling the user to utilize the
collected empirical information (even if limited in amount) in a reliable and
validated way. Being a hybrid between probabilistic and deterministic models,
clouds can provide risk analysis using tools from optimization, in particular
global optimization, and constraint satisfaction techniques. The numerical
techniques for solving such problems have recently become much more reliable
and powerful and allow one to compute bounds for the expected values of
any multivariate functions of design processes and also for probabilities of
qualitative statements involving design variables [NFD+ 07].

12.5 Summary
The aim of this chapter is to give the reader an overview of some of the
research carried out within the international space community on artificial
intelligence. Having identified artificial intelligence as one of the enabling
technologies for the achievement of the various short- and long-term goals
of the international space agencies, we believe that a synergic effort of
scientists from both fields is required to effectively tackle the numerous open
issues and challenges in this area. In more recent years, we have observed
a growing number of researchers getting interested in the benefits of using
artificial intelligence methods for space applications. These applications go
beyond the more classical automated planning and scheduling field and include
different mission phases from conceiving the preliminary design to the mission
operation phase.

References
[ADL98] J.A. Aguilar, A.B. Dawdy, and G.W. Law. The Aerospace Corpora-
tion’s concept design center. In 8th Annual International Symposium of
the International Council on Systems Engineering, volume 2. INCOSE
Technical Working Group Papers, 1998. Published in CD-ROM.
[AF98] I.P. Androulakis and C.A. Floudas. Distributed branch and bound algo-
rithms in global optimization. In P.M. Pardalos, editor, IMA Volumes
 12 Artificial Intelligence for Space Applications 249

in Mathematics and Its Applications, volume 106, pages 1–36. Springer-

Verlag, New York, 1998.
[AFA+ 04] V. Amata, G. Fasano, L. Arcaro, F. Della Croce, M.F. Norese,
S. Palamara, R. Tadei, and F. Fragnelli. Multidisciplinary optimisa-
tion in mission analysis and design process. GSP programme ref. GSP
03/N16, contract number 17828/03/NL/MV, European Space Agency,
Noordwijk, 2004.
[BMN+ 05] V.M. Becerra, D.R. Myatt, S.J. Nasuto, J.M. Bishop, and D. Izzo. An
efficient pruning technique for the global optimisation of multiple grav-
ity assist trajectories. In I. Garcia, L.G. Casado, E.M.T. Hendrix, and
B. Toth, editors, Proceedings of the International Workshop on Global
Optimization, pages 39–45, Almeria, Spain, April 2005.
[BMO99] M. Bandecchi, S. Melton, and F. Ongaro. Concurrent engineering applied
to space mission assessment and design. ESA Bulletin, 99, September
1999. Available at http://esapub.esrin.esa.it/bulletin/bullet99.
htm.
[Bro91] R.A. Brooks. Intelligence without reason. In J. Mylopoulos and
R. Reiter, editors, Proceedings of the 12th International Joint Confer-
ence on Artificial Intelligence, pages 569–595, Sydney, Australia, August
1991. Morgan Kaufmann, San Mateo, CA.
[BS02] V. Belton and T.J. Stewart. Multiple Criteria Decision Analysis: An
Integrated Approach. Kluwer Academic Publishers, Dordrecht, 2002.
[BT07] G. Bonnet and C. Tessier. On-board cooperation for satellite swarms. In
D. Girimonte, D. Izzo, T. Vinko, S. Chien, and V. Kreinovich, editors,
Workshop on Artificial Intelligence for Space Applications at IJCAI’07,
pages 26–27, Hyderabad, India, January 2007. IJCAI.
[Bus04] G.W. Bush. A renewed spirit of discovery: the President’s vision for U.S.
space exploration. 2004.
[CCD+ 05] S. Chien, B. Cichy, A. Davies, D. Tran, G. Rabideau, R. Castano,
R. Sherwood, D. Mandl, S. Frye, S. Shulman, J. Jones, and S. Grosvenor.
An autonomous Earth-observing sensorweb. IEEE Intelligent Systems,
20(3):16–24, May–June 2005.
[CCV07] N. Croisard, M. Ceriotti, and M. Vasile. Uncertainty modelling in reliable
preliminary space mission design. In D. Girimonte, D. Izzo, T. Vinko,
S. Chien, and V. Kreinovich, editors, Workshop on Artificial Intelligence
for Space Applications at IJCAI’07, pages 22–23, Hyderabad, India, Jan-
uary 2007. IJCAI.
[Cla45] A.C. Clarke. Extra-terrestrial relays. Wireless World, LI(10):305–308,
October 1945.
[CST+ 04] S. Chien, R. Sherwood, D. Tran, B. Cichy, G. Rabideau, R. Castano,
A. Davies, R. Lee, D. Mandl, S. Frye, B. Trout, J. Hengemihle,
J. D’Agostino, S. Shulman, S. Ungar, T. Brakke, D. Boyer, J. Van Gaas-
beck, R. Greeley, T. Doggett, V. Baker, J. Dohm, and F. Ip. The EO-1
autonomous science agent. In Proceedings of the 3rd International Joint
Conference on Autonomous Agents and Multiagent Systems, pages 420–
427, New York, New York, 19–23 July 2004. IEEE Computer Society,
Washington, DC.
[DP05] D. Dubois and H. Prade. Interval-valued fuzzy sets, possibility theory
and imprecise probability. In E. Montseny and P. Sobrevilla, editors,
250 Daniela Girimonte and Dario Izzo

Proceedings of the International Conference in Fuzzy Logic and Technol-

ogy, pages 314–319, Barcelona, Spain, 8–10 September 2005. Universitat
Politènica de Catalunya.
[DRIV05] P. Di Lizia, G. Radice, D. Izzo, and M. Vasile. On the solution of in-
terplanetary trajectory design problems by global optimisation methods.
In I. Garcia, L.G. Casado, E.M.T. Hendrix, and B. Toth, xeditors, Pro-
ceedings of the International Workshop on Global Optimization, pages
159–164, Almeria, Spain, April 2005. Available at http://www.esa.int/
gsp/ACT/publications/pub-mad.htm.
[DVC05] S. Damiani, G. Verfaillie, and M.C. Charmeau. An Earth watching satel-
lite constellation: how to manage a team of watching agents with limited
communications. In Proceedings of the 4th International Joint Confer-
ence on Autonomous Agents and Multiagent Systems, AAMAS, pages
455–462, Utrecht, Netherlands, 25–29 July 2005. ACM Press, New York,
NY.
[ETR+ 03] D. Escorial, I.F. Tourne, F.J. Reina, J. Gonzalo, and B. Garrido. Fuego:
a dedicated constellation of small satellites to detect and monitor forest
fires. Acta Astronautica, 52(9–12):765–775, 2003.
[FCM+ 97] A. Fukunaga, S. Chien, D. Mutz, R.L. Sherwood, and A.D. Stechert. Au-
tomating the process of optimization in spacecraft design. In Proceedings
of the IEEE Aerospace Conference, volume 4, pages 411–427, Snowmass
at Aspen, CO, 1–8 February 1997. IEEE, New York.
[Fey86] R. Feynman. Report of the presidential commission on the space shuttle
Challenger accident. Volume 2 (appendix f), U.S. Government Printing
Office, Washington, DC, 1986.
[FGL+ 03] L. Fusco, P. Goncalves, J. Linford, M. Fulcoli, A. Terracina, and
G. D’Acunzo. ESA Bulletin, 114:86–90, 2003.
[Gaz05] V. Gazi. Swarm aggregations using artificial potentials and sliding mode
control. IEEE Transactions on Robotics, 21(6):1208–1214, December
2005.
[Geh03] H.W. Gehman. The Columbia accident investigation board report. Tech-
nical report, U.S. Government Printing Office, Washington, DC, 2003.
[GIB06] D. Girimonte, D. Izzo, and L. Bergamin. Reasoning under an uncertain
thermal state. In Proceedings of the 57th International Astronautical
Congress, Valencia, Spain, 2–6 October 2006. International Astronautical
Federation.
[GP02] C.; Gagne and M. Parizeau. A new versatile C++ framework for evo-
lutionary computation. In W.B. Langdon, E. Cantú-Paz, K. Mathias,
R. Roy, D. Davis, R. Poli, K. Balakrishnan, V. Honavar, G. Rudolph,
J. Wegener, L. Bull, M.A. Potter, A.C. Schultz, J.F. Miller, E. Burke,
and N. Jonoska, editors, Proceeding of Genetic and Evolutionary Compu-
tation Conference, page 888, New York, NY, April 2002. Morgan Kauf-
mann, San Francisco, CA.
[HGLL06] G.S. Hornby, A. Globus, D.S. Linden, and J.D. Lohn. Automated an-
tenna design with evolutionary algorithms. In Proceedings of the AIAA
Space Conference, San Jose, California, 19–21 September 2006. Available
at http://ase.arc.nasa.gov/publications.
[HSC+ 04] S.C. Hayden, A.J. Sweet, S.E. Christa, D. Tran, and S. Shulman. Ad-
vanced diagnostic system on Earth Observing One. In Proceedings of the
 12 Artificial Intelligence for Space Applications 251

AIAA Space Conference, 2004. Available at http://ase.arc.nasa.gov/

publications.
[IBM+ 07] D. Izzo, V.M. Becerra, D.R. Myatt, S.J. Nasuto, and J.M. Bishop. Search
space pruning and global optimisation of multiple gravity assist space-
craft trajectories. Journal of Global Optimisation, 38(2):283–296, June
2007.
[IM05] D. Izzo and M.Cs. Markot. A distributed global optimisation environ-
ment for the European Space Agency internal network. In I. Garcia,
L.G. Casado, E.M.T. Hendrix, and B. Toth, editors, Proceedings of
the International Workshop on Global Optimization, pages 133–140,
Almeria, Spain, April 2005. Available at http://www.esa.int/gsp/ACT/
publications/pub-inf.htm.
[IMN05] D. Izzo, M.Cs. Markot, and I. Nann. A distributed global optimiser ap-
plied to the design of a constellation performing radio-occultation mea-
surments. In Univelt Inc., editor, Proceedings of the 2005 AAS/AIAA
Space Flight Mechanics Conference, volume 121, pages 739–748, Tampa,
Florida, 2005. AAS Publications Office. Paper AAS 05–150.
[IP07] D. Izzo and L. Pettazzi. Autonomous and distributed motion planning for
satellite swarm. Journal of Guidance Control and Dynamic, 30(2):449–
459, 2007.
[Ive04] I.L. Iverson. Inductive system health monitoring. In H.R. Arabnia and
Y. Mun, editors, Proceedings of the International Conference on Artificial
Intelligence IC-AI’04, volume 2, pages 605–611, Las Vegas, Nevada, 21–
24 June 2004. CSREA Press.
[IVH06] Integrated vehicle health management (IVHM), March 2006. Available
at http://www.nasa.gov/centers/ames/research/.
[KE95] J. Kennedy and R. Eberhart. Particle swarm optimization. In Proceedings
of the IEEE International Conference on Neural Networks, volume 4,
pages 1942–1948, Perth, Australia, 1995. IEEE, New York.
[KEM+ 01] I. Kassabalidis, M.A. El-Sharkawi, R.J. Marks II, P. Arabshahi, and
A. Gray. Swarm intelligence for routing in communication networks.
In Proceedings of the IEEE Global Telecommunications Conference
GLOBECOM, volume 6, pages 3613–3617, San Antonio, TX, 25–29
November 2001. Available at http://ieeexplore.ieee.org/xpls/abs
all.jsp?arnumber=966355.
[Kha86] O. Khatib. Real-time obstacle avoidance for manipulators and mobile
robots. The International Journal of Robotics Research, 5(1):90–98, 1986.
[KMSR03] G. Karpati, J. Martin, M. Steiner, and K. Reinhardt. The Integrated
Mission Design Center (IMDC) at NASA Goddard Space Flight Center.
In Proceedings of the IEEE Aerospace Conference, volume 8, pages 3657–
3667. IEEE, New York, 8–15 March 2003.
[LF02] M. Lavagna and A.E. Finzi. A multi-attribute decision-making approach
toward space system design automation through a fuzzy logic-based an-
alytic hierarchical process. In T. Hendtlass and M. Ali, editors, Proceed-
ings of the 15th International Conference on Industrial and Engineering.
Applications of Artificial Intelligence and Expert Systems, volume 2358
of Lecture Notes In Computer Science, pages 596–606. Springer-Verlag,
London, 17–20 June 2002.
252 Daniela Girimonte and Dario Izzo

[LYB03] J. Lawton, B. Young, and R. Beard. A decentralized approach to

elementary formation maneuvers. IEEE Transactions on Robotics and
Automation, 17(6):933–941, 2003.
[MBNB04] D.R. Myatt, V.M. Becerra, S.J. Nasuto, and J.M. Bishop. Advanced
global optimisation tools for mission analysis and design. Technical Re-
port 03-4101, European Space Agency, Noordwijk, 2004. Available at
http://www.esa.int/gsp/ACT/ariadna/completed\ studies.htm.
[McI95] C. McInnes. Autonomous rendezvous using artificial potential functions.
Journal of Guidance Control and Dynamics, 18(2):237–241, 1995.
[McQ97] F. McQuade. Autonomous control for on-orbit assembly using artificial
potential functions, 1997. PhD thesis, Faculty of Engineering, University
of Glasgow, Scotland.
[MO03] P. Messina and F. Ongaro. Aurora: the European space exploration
programme. ESA Bulletin, 115:34–39, 2003.
[NBB04] S. Narasimhan, L. Brownston, and D. Burrows. Explanation constraint
programming for model-based diagnosis of engineered systems. In Pro-
ceedings of the IEEE Aerospace Conference, volume 5, pages 3495–3501,
Big Sky, Montana, 6–13 March 2004. IEEE, New York.
[NDB04] S. Narasimhan, R. Dearden, and E. Benazera. Combining particle filters
and consistency-based approaches for monitoring and diagnosis of sto-
chastic hybrid systems. In Proceedings of the 15th International Work-
shop on Principles of Diagnosis DX’04, Carcassonne, France, 23–25 June
2004. Available at http://www.laas.fr/DX04/.
[Neu04] A. Neumaier. Clouds, fuzzy sets and probability intervals. Reliable Com-
puting, 10:249–272, 2004.
[NFD+ 07] A. Neumaier, M. Fuchs, E. Dolejsi, T. Csendes, J. Dombi, B. Banhelyi,
and Z. Gera. Application of clouds for modeling uncertainties in ro-
bust space system design. ACT Ariadna Research ACT-RPT-05-5201,
European Space Agency, Noordwijk, 2007. Available at http://www.
esa.int/gsp/ACT/ariadna/completed\ studies.htm.
[PF93] M. Pate-Cornell and P. Fischbeck. Probabilistic risk analysis and risk
based priority scale for the tiles of the space shuttle. Reliability Engi-
neering and System Safety, 40(3):221–238, 1993.
[PIT06] L. Pettazzi, D. Izzo, and S. Theil. Swarm navigation and reconfiguration
using electrostatic forces. In Proceedings of the 7th International Con-
ference on Dynamics and Control of Systems and Structures in Space,
pages 257–268, Greenwich, UK, 16-20 July 2006. Cranfield University
Press, London.
[RB04] W. Ren and R.W. Beard. A decentralized scheme for spacecraft formation
flying via the virtual structure approach. AIAA Journal of Guidance,
Control and Dynamics, 27(1):73–82, January 2004.
[RC96] G.A. Rauwolf and V.L. Coverstone-Carroll. Near optimal low-thrust
orbit transfers generated by a genetic algorithm. Journal of Spacecraft
and Rockets, 33(6):859–862, 1996.
[Roy96] B. Roy. Multicriteria Methodology for Decision Aiding. Kluwer Academic
Publishers, Dordrecht, 1996.
[Sch05] M. Schwabacher. Machine learning for rocket propulsion health monitor-
ing. SAE Transactions, 114(1):1192–1197, 2005.
 12 Artificial Intelligence for Space Applications 253

[SD92] G. Schoner and M. Dose. A dynamics systems approach to task level

systems integration used to plan and control autonomous vehicle motion.
Robotics and Autonomous Systems, 10(4):253–267, 1992.
[TB99] G. Theraulaz and E. Bonebeau. A brief history of stigmergy. Artificial
Life, 5(2):97–116, 1999.
[Thu05] D.P. Thunnissen. Propagating and mitigating uncertainty in the design of
complex multidisciplinary systems, 2005. PhD thesis, California Institute
of Technology, Pasadena.
[VIP07] T. Vinko, D. Izzo, and F. Pinna. Learning the best combination of solvers
in a distributed global optimisation environment. Technical Report ACT-
RPT-TV-5200-LBCSDGOE, ESA Advanced Concepts Team, Noordwijk,
2007.
[VLFD05] E. Vecchio, B. Lazzerini, S. Foley, and A. Donati. Spacecraft fault analy-
sis using data mining techniques. In B. Battrick, editor, Proceedings of
the 8th International Symposium on Artificial Intelligence, Robotics and
Automation in Space, München, Germany, 5–8 September 2005. Pub-
lished in CDROM.
[WL99] J.R. Wertz and W.J. Larson. Space Mission Analysis and Design. Micro-
cosm Press, Torrance, CA, 3rd edition, 1999. Space Technology Library.
[WN96] B.C. Williams and P. Nayak. A model-based approach to reactive self-
configuring systems. In Proceedings of the 13th National Conference
on Artificial Intelligence, volume 2, pages 971–978, Portland, OR, 4–8
August 1996. AAAI Press/The MIT Press.

View publication stats