ANZIAM J.

46 (E) ppC89–C101, 2005 C89

Comparing genetic algorithms and particle

swarm optimisation for an inverse problem
exercise

C. R. Mouser∗ S. A. Dunn∗

(Received 26 October 2004, revised 1 February 2005)

Abstract

We describe the performance of two population based search al-

gorithms (genetic algorithms and particle swarm optimisation) when
applied to the optimisation of a structural dynamics model. A signifi-
cant difficulty arises when trying to compare the performance of such
algorithms. For the two algorithms to perform at their best, several
properties (for example, population size and mutation rate) need to
be set. The performance of the algorithms can be highly sensitive to
the choice of these parameters, and the optimisation of these leads
to a search in a multi-dimensional space. This work describes how
a genetic algorithm optimises the properties of a genetic algorithm
and a particle swarm optimisation in order to produce algorithms
∗
Defence Science & Technology Organisation, Melbourne, Australia.
mailto:[email protected]
See http://anziamj.austms.org.au/V46/CTAC2004/Mous for this article, c Aus-
tral. Mathematical Soc. 2005. Published March 21, 2005. ISSN 1446-8735
ANZIAM J. 46 (E) ppC89–C101, 2005 C90

that are optimally tuned to the particular problem being solved. The
two methods are rigorously compared. This problem is implemented
on a distributed computing facility spread across the Defence Science
& Technology Organisation’s network across four cities in south-east
Australia.

Contents
1 Introduction C90

2 Structural dynamics model C92

3 Optimising the structural dynamics model C93

3.1 Genetic algorithm . . . . . . . . . . . . . . . . . . . . . . . C94
3.2 Particle Swarm Optimisation . . . . . . . . . . . . . . . . . C95

4 Optimising the optimiser C95

5 Distributed computing C98

6 Results C99

7 Conclusion C101

References C101

1 Introduction

There is a long history of development and updating of dynamic finite element

models based on measured experimental data. These models are created
at the time of design of the aircraft and a significant amount of time is
1 Introduction C91

expended in developing them to the fidelity required. Because these models

are used in assessing the flight-worthiness of the aircraft, a great importance
is placed on them. Ground Vibration Tests (gvts) are then conducted on
the aircraft when it is built and the model is validated against experiment.
Typically, the agreement between the model and reality is poor. A process
of updating the model is then embarked upon so as to make the model a
better predictor of the actual dynamic response of the aircraft. Typically,
this involves the adjustment of the physical parameters in the model (such
as moduli of elasticity and joint stiffness representation). This process still
leaves all the assumptions that were inherent in the modelling process and
the final model is typically a compromise. Also, the number of parameters
that can be adjusted are typically significantly larger than the information
contained in the experimental data.

Alternatively, the finite element model can be used as a starting point

to produce mass and stiffness matrices that approximate the structure. In-
dividual elements of these matrices are altered to create a model that most
accurately represents the structure under consideration. This process suffers
from a loss of physical causality between the model and the structure and can
lead to such physical impossibilities as negative mass. It is difficult to have
a very high degree of confidence in the predictions of such a model following
physical modifications to a structure

Another approach, that considered here, is to build a model based on

measured experimental data [1]. A relatively simple model is constructed
based on the physics of the structure. This is constructed from basic finite
elements such as lumped masses, beams that can support bending and tor-
sion, springs and dampers. The properties of all these elements are assumed
to be unknown and an optimisation problem is constructed to determine the
values of the model parameters based on the measured data. Even though
greatly simplified, the model complexity is typically such that the number of
unknown parameters are in the hundreds. Traditional optimisation routines
have great difficulty in solving problems of this magnitude. However, the
1 Introduction C92

Model
Frequency Response, χ

εi,2 Experimental
εi,1 εi,N-1

εi,N

Frequency, ω

Figure 1: Sample frequency response function

field of evolutionary optimisation has given rise to a new class of optimisers

to apply to such a problem. This paper will discuss two such optimisation
routines, namely genetic algorithms and particle swarm optimisation.

2 Structural dynamics model

To investigate the suitability of using genetic algorithms or particle swarm

optimisation for solving the class of problems being discussed here, a con-
ceptually simple problem was contrived. This consists of a 10 degree-of-
freedom mass/spring/damper system, for which the values of mass, stiffness
and damping are known. From this system, theoretical frequency response
functions were calculated to simulate measured experimental data. We then
assume that the parameters (mass, stiffness, damping) of the model are un-
known. Hence we construct an optimisation problem to minimise the differ-
ence between the “measured” frequency response functions and the calculated
2 Structural dynamics model C93

frequency response functions. To control the difficulty of the problem, the

number of unknown masses, springs and dampers is varied to give between
two and thirty unknowns. Figure 1 gives an example frequency response
function, showing both the measured and experimental curves.

Based on Figure 1, a cost function is developed. Given a vector of un-

knowns, η, representing the unknown physical properties of the system, the
objective is to minimise the error between the model predictions, χ (η) and
the measurements χ0 (η) for the n freedoms at the N selected comparison
frequencies:
N X n
0
X
min ( (η)) = |χi,j (η) − χi,j (η)| . (1)
j=1 i=1

The model predictions are given by the solution to the linear equation

−ω 2 [M (η)] + iω [c (η)] + [K (η)] χ (η) = φ ,

(2)

where M is the mass matrix, K is the stiffness matrix, c is the damping ma-
trix and φ is a vector of amplitudes of a sinusoidal input force at frequency ω.
Note that M , K and c are functions of the unknowns, η. This implies that
the structure of these matrices is constrained by the physics of the original
mass/spring/damping system matrix structure and within the optimisation
procedure this structure is maintained. This ensures that the final model,
while it may not be the best mathematical model for the data, is the best
model for the data without compromising the physics of the problem. Con-
sequently, the model maintains its predictive capability.

3 Optimising the structural dynamics model

Based on the structural dynamics model given above, the aim of this work
is to compare the performance of two optimisation algorithms, when applied
3 Optimising the structural dynamics model C94

to this problem. The difficulty with the application of the optimisation al-
gorithms is that they need to be tuned. For each algorithm there are several
parameters to vary to control the algorithm, the settings of these parameters
can have a significant impact on the performance of the optimisation routine.
If the parameters are set poorly, the algorithm will either converge slowly, or
may not converge at all. There are general guidelines given as to how these
parameters should be set, but experience has shown that the optimal value
of these parameters is very problem specific.

The basis for the comparison between genetic algorithms and the particle
swarm optimisation is the number of cost function evaluations required to
get the solution within a specified tolerance. A cost function calculates the
frequency response function from the generated structural dynamic model
and compares this to the simulated-measured frequency response function,
returning a fitness value for the solution.

3.1 Genetic algorithm

Genetic algorithms are applied to structural dynamic problems at dsto [1].

Briefly, the genetic algorithm consists of a population of individuals, rep-
resented by a binary string, that encode the unknowns of the optimisation
problem. At each generation of the solution process, individuals are selected
to breed and produce a new population of individuals. For this applica-
tion, tournament selection, with uniform crossover and mutation was used.
A niching strategy has also been adopted, whereby after a certain number
of generations have evolved, the best solution(s) found are used to form a
new population of individuals, along with randomly chosen individuals. The
number of individuals to carry through, along with the number of genera-
tions to evolve before restarting are parameters that can be varied to control
the performance of the algorithm.
3 Optimising the structural dynamics model C95

3.2 Particle Swarm Optimisation

Particle Swarm optimisation is a relatively new optimisation technique [2].

It is inspired by the swarming behaviour of birds and insects. A population
of particles exists in the n-dimensional search space that the optimisation
problem lives in. Each particle has a certain amount of knowledge, and
will move about the search space based on this knowledge. The particle
has some inertia attributed to it and so will continue to have a component
of motion in the direction it is moving. It also knows where in the search
space the best solution it has encountered is, and finally it knows where the
best solution any of the other particles has found is. The particle will then
modify its direction such that it has additional components towards its own
best position and towards the overall best position. This will provide some
form of convergence to the search, while providing a degree of randomness
to promote a wide coverage of the search space. For the ith particle, in the
dth dimension, the algorithm is
vi,d (n + 1) = wvi,d (n) + c1 r1,d (pi,d − xi,d (n)) + c2 r2,d (pg,d − xi,d (n)) ,
(3)
xi,d (n + 1) = xi,d (n) + vi,d (n + 1) ,
where, vi,d represents the velocity of the particle at time step (n + 1) and xi,d
represents the position of the particle. The initial position and velocity of
the particles are chosen randomly. The values of w, c1 and c2 are fixed
for a given run of the particle swarm optimisation and are set by the outer
genetic algorithm. r1,d and r2,d are uniformly distributed random numbers.
pi,d and pg,d are the best position the ith particle has currently found and
the best position found by any particle respectively.

4 Optimising the optimiser

In order to accurately compare the performance of the two optimisation algo-

rithms, the algorithms have to be performing optimally. In order to determine
4 Optimising the optimiser C96

the optimum properties for the optimisation routines, a genetic algorithm is

used here to optimise the optimiser. This outer genetic algorithm attempts
to find the inner genetic algorithm or particle swarm optimisation properties
that will optimise the structural dynamics model in the minimum time (as
measured by the number of cost function evaluations). The properties of the
outer genetic algorithm are fixed. The properties being optimised by this
outer genetic algorithm are:

Genetic algorithm Particle Swarm Optimisation

• population size • number of particles
• total number of ga runs • inertia (w)
• proportion niched • own-best weighting (c1 )
• pre-niche mutation rate • global-best weighting (c2 )
• post-niche mutation rate • maximum allowable vi,d

For the outer genetic algorithm to operate, a cost function needs to be

defined. This cost function needs to be able to return a measure of the success
of the optimisation routine applied to the structural dynamics model. To be
considered successful, an optimisation routine needs to meet several criteria:

1. It must be able to find an acceptable solution;

2. It should do this as quickly as possible (that is, with the minimum

number of cost function evaluations); and

3. For any given run, it should have a high probability of achieving the
first two goals.

This last point is an important consideration. Due to the non-deterministic

nature of the optimisation routines (that is, they have a random starting
point, and their evolution has an element of randomness present) one set
of parameters for an optimiser may find a solution only occasionally. Such
4 Optimising the optimiser C97

a set of parameters cannot be considered to define a good optimiser as the

likelihood of the optimiser converging is too low. The defined cost function
needs to incorporate all of these criteria.

For each run of the inner optimisers, a fixed number of cost function eval-
uations is performed (Ncfe ). For each set of parameters, the inner optimiser
is run νrun times. This aids in providing a measure of the repeatability of
the optimiser. The number of runs with an acceptable cost function (Cacc )
must be greater than νrun /2, otherwise a strong penalty is applied. This
ensures that optimisers that are not repeatable are punished and the outer
genetic algorithm should move away from these solutions. Based on these
considerations, the cost function for the outer genetic algorithm is described
as follows:

For a given set of genetic algorithm or particle swarm optimisa-

tion parameters from the outer genetic algorithm, run the inner
optimiser 10 times. For the inner genetic algorithm or particle
swarm optimisation, increment the number of cost function eval-
uations of the inner optimiser (ncfe ) until cfinner < Cacc or until
ncfe = Ncfe . Note that if cfinner < Cacc , the optimiser contin-
ues until ncfe = Ncfe but ncfe is not incremented. The following
algorithm was then used for the final cost function:

if νsuccessful ≥ νrun /2
X
cfouter = ncfe /νsuccessful
else
X
cfouter = 107 cfinner

cfinner and cfouter are the value of the cost function for the in-
ner and outer optimisers respectively. νsuccessful is the number of
successful runs of the inner optimiser.
4 Optimising the optimiser C98

Note that this cost function will reward parameters that show they can
repetitively achieve an acceptable solution and punish those that do not.
After the outer genetic algorithm has run through a prescribed number of
iterations, the solutions that look promising are then run 100 times to more
rigorously assess their performance. The average number of cost function
evaluations from these 100 runs is then used to form the basis for comparison
between the particle swarm optimisation and the genetic algorithm.

As each member of the outer population performs 10 independent op-

timisations, and the results from all the members only need to be merged
once all the outer population members have completed their calculations, the
problem is conceptually easy to run as a parallel algorithm.

5 Distributed computing

Problems of the magnitude being considered here are difficult, if not impos-
sible, to solve on current generation desktop pcs. However, dsto has a large
number of desktop pcs that sit idle during nights, weekends and holidays.
The numerical processing ability of 100s of desktop pcs is applied to the
optimisation. To this end, the pcs at dsto are used out-of-hours to perform
the optimisation [3]. The system is constructed with a dedicated server pro-
viding data storage capability. Each evening, participating pcs are rebooted
from a floppy disk and automatically connect to the server where they are
allocated a job. Each computer then processes its allocated job and, when it
is completed, submit the answer and be allocated a new job.

When an individual pc (referred to as a drone) comes on-line, it checks

the job control file to determine which jobs are still to be completed. It
then obtains the parameters to use for the inner optimisation routine and
will run this job to completion. When finished, the drone will write the
results back into a file on the server. The results written are the number of
function evaluations required to optimise the structural model and the value
5 Distributed computing C99

of the cost function after performing the maximum allowed number of cost
function evaluations. The drone will then request another job. If there is
nothing remaining to process, the drone will take all the results currently
available and use these to evaluate the fitness function for each individual
of the outer genetic algorithm and advance the outer ga by one generation.
This will then create a new array that needs to be filled and jobs that can
be allocated to new drones.

The process has been setup such that when a drone logs on, it will be allo-
cated a job, and if a drone “dies” before completing its allocated task (either
the computer is rebooted for normal operation at the start of a working day,
or there is a network or software failure) the job that drone was processing
will be allocated to a new drone. Also, the nature of the system means that
not all computers attached to the network are equal in terms of performance.
If for one or more given generations of the outer genetic algorithm, the pro-
cess is waiting for slow drones to return a result, another drone or drones
will start processing the outstanding jobs with the hope of completing the
job quicker. If when a drone returns a job it finds that a result has already
been recorded, it will discard its result and obtain a new job. In this way, the
system is robust against drones joining and leaving the network, hardware
or software failures and varying drone performance capabilities. As long as
the central server and at least one drone exist, the processing can continue.

6 Results

Current results indicate that the particle swarm optimisation outperforms

the genetic algorithm for problems of the nature being considered here. See
in Figure 2 that the number of cost function evaluations can be an order
of magnitude lower for a particle swarm algorithm, when compared to the
genetic algorithm. Also shown in this figure are the results from a simplex
optimisation algorithm. The poor results of this algorithm highlight the
6 Results C100

1E+009
1.8n Simplex
N=1.4e PSO
1E+008 GA
number of cost function evaluations, N

1E+007

N=14n4.8
1E+006 N=2.4n4.3

1E+005

1E+004

1E+003

1E+002

1E+001
0 5 10 15 20 25 30
number of unknowns, n

Figure 2: Number of cost function evaluations required for a given problem

size

benefits of a population based optimisation, such as a genetic algorithm or

particle swarm optimisation.

Note that the genetic algorithm processing has not yet advanced past
11 unknowns. Work is still underway to progress this line as far as possible.
However, see that a genetic algorithm with 11 unknowns is taking a similar
number of cost function evaluations to solve as a particle swarm optimisation
with 25 unknowns. The amount of computing power required to advance the
genetic algorithm line is significant and progress is slow.

As the number of unknowns increased, the optimal properties of the par-

6 Results C101

ticle swarm optimisation changed in a much more predictable manner than

those of the genetic algorithm. The characterisation of this change is the
subject of further investigation and will be covered in a subsequent paper.

7 Conclusion

We have shown that it is possible to use evolutionary optimisation algorithms

to create an optimum structural dynamics model. We also found that the
selection of the optimisation algorithm has a significant effect on the suit-
ability of the final model. For the two optimisation algorithms considered
here, the particle swarm optimisation algorithm significantly outperformed
the genetic algorithm. Also, the particle swarm optimisation algorithm is
much easier to configure than the genetic algorithm and is more likely to
produce an acceptable model.

References

[1] S. A. Dunn, Optimised Structural Dynamic Aircraft Finite Element

Models Involving Mass, Stiffness and Damping Elements. International
Forum on Aeroelasticity and Structural Dynamics, Madrid, Spain, June
2001, pp.387–396 C91, C94

[2] J. Kennedy and R. C. Eberhaurt, Particle Swarm Optimization. Proc.

IEEE Int. Conf. Neural Networks, Perth, Australia Nov. 1995,
pp.1942–1948 C95

[3] S. Dunn, S. Peucker and J. Perry, Genetic Algorithm Optimisation of

Mathematical Models Using Distributed Computing. Applied
Intelligence, in-press C98