Adaptive Weighted Particle Swarm Optimisation

for Multi-objective Optimal Design of Alloy Steels

Mahdi Mahfouf, Min-You Chen, and Derek Arthur Linkens

Institute for Microstructure and Mechanical Properties Engineering,

The University of Sheffield (IMMPETUS)
Department of Automatic Control and Systems Engineering,
The University of Sheffield, Mappin Street, Sheffield S1 3JD, UK
{M.Mahfouf,minyou.chen,d.linkens}@shef.ac.uk

Abstract. In this paper, a modified Particle Swarm Optimisation (PSO) algo-

rithm is presented to improve the performance of multi-objective optimisation.
The PSO algorithm search capabilities are enhanced via the inclusion of the
adaptive inertia weight and acceleration factor. In addition, a weighted aggrega-
tion function has been introduced within the algorithm to guide the selection of
the personal and global bests, together with a non-dominated sorting algorithm
to select the particles from one iteration to another. The proposed algorithm has
been successfully applied to a series of well-known benchmark functions as
well as to the multi-objective optimal design of alloy steels, which aims at de-
termining the optimal heat treatment regimes and the required weight percent-
ages for the chemical composites in order to obtain the pre-defined mechanical
properties of the material. The results have shown that the algorithm can locate
the constrained optimal design with a very good accuracy.

1 Introduction
Multi-objective optimisation is becoming more and more the focus of active research
for many real-world problems most of which are indeed ‘multi-objective’ in nature. In
many scientific and engineering environments, it is not uncommon to face design
challenges when there are several criteria or design objectives which need to be met
concomitantly. If these objectives happen to conflict with each other, then the prob-
lem becomes equivalent to finding the best possible design(s) that satisfy the compet-
ing objectives under predefined trade-off scenarios. In the steel industry in particular,
optimal metal design represents a challenging multi-objective optimisation problem,
which consists of finding the optimal processing parameters and the corresponding
chemical compositions to obtain certain pre-defined mechanical properties. Particle
Swarm Optimisation (PSO) is a relatively new technique for finding optimal regions
of complex search spaces through the interaction of individuals in a population of
particles. J. Kennedy and C. Eberhart originally introduced the technique in 1995 [1].
Unlike evolutionary algorithms, which are based on the principle of survival of the
fittest, PSO is motivated by the simulation of the social behaviour of flocks. As Ken-
nedy stated [2], the algorithm is based on a metaphor of social interaction, searches a
space by adjusting the trajectories of individual vectors, called “particles” as they are
conceptualised as moving points in the multidimensional space. The individual parti-
cles evaluate their positions, at each iteration, relative to a goal (fitness). They are

X. Yao et al. (Eds.): PPSN VIII, LNCS 3242, pp. 762–771, 2004.
© Springer-Verlag Berlin Heidelberg 2004
 Adaptive Weighted Particle Swarm Optimisation for Multi-objective Optimal Design 763

drawn stochastically towards the positions of their own previous best performance
and the best previous performance of their companions. The PSO algorithm has been
shown to be a successful optimiser over a wide range of functions [3]-[6], and at-
tracted wide attention from several scientific and engineering communities. This pa-
per represents an enhancement of the original PSO algorithm by introducing adaptive
weighting and a weighted aggregation function for performance evaluation. The
modified algorithm has been successfully tested on a set of difficult functions and
subsequently applied to multi-objective optimal design of alloy steels.
The remaining parts of the paper are organised as follows: Section 2 presents the
adaptive weighted non-dominated sorting PSO algorithm. Section 3 will show the
results of simulations using the proposed algorithm to solve the multi-objective opti-
misation test problems ZDT1~ZDT4. This section will also conduct a comparative
study between the proposed algorithm and other commonly recognised effective
multi-objective optimisation algorithms. The analysis and results relating to the opti-
mal design of alloy steels will be given in Section 4. Finally, concluding remarks will
be given in Section 5.

2 The Modified Particle Swarm Optimisation Algorithm

The particle swarm algorithm imitates human social behaviour. Individuals interact
with one another while learning from their own experience(s), and gradually, the
population members move into better regions of the problem space. The algorithm is
incredibly simple – it can be described in one straightforward formula – but is able to
circumvent many of the obstacles that optimisation problems usually present, includ-
ing those associated with Genetic Algorithms (GA).
The PSO algorithm defines each particle as a potential solution to a problem in a
dimensional space, with the particle I, represented by the following vector:
xi=(xi1, xi2, …, xid), i=1, 2, … N
Where d is the search dimension and N is the number of particles in the population.
The original formula, which was developed by Kennedy and Eberhart, was later
improved by Shi and Eberhart by introducing an inertia weight w [6]. During each
iteration, the particle’s position is modified according to the following equations:
vi(t) = w vi(t-1)+c1r1(pi- xi(t-1))+ c2r2(pg- xi(t-1)) (1)
xi(t) = vi(t)+ xi(t-1) (2)
where vi is the velocity, w is the inertia weight, c1 and c2 are positive constants, and r1
and r2 are random numbers obtained from a uniform random distribution function in
the interval [0, 1]. The parameter pi represents the best previous position of the ith
particle and pg denotes the best particle among all the particles in the population.
The inertia weight w plays the role of balancing the global and local searches and
its value may vary during the optimisation process. A large inertia weight encourages
a global search while a small value pursues a local search. In [6], Shi and Eberhart
suggested to change the inertia weight linearly from 1 to 0.4 to restrict the global
search ability of the PSO algorithm at the end of a run. They also proposed a ‘fuzzy
adaptation’ of the inertia weight [7] due to the fact that a linearly-decreasing weight
764 Mahdi Mahfouf, Min-You Chen, and Derek Arthur Linkens

would not be adequate to improve the performance of the PSO due to its non-linear
nature.
To improve the performance of the PSO for multi-objective optimisation problems,
we hereby propose an Adaptive Weighted PSO (AWPSO) algorithm, in which the
velocity in Equation (1) is modified as follows:
vi(t+1) = w vi(t)+α [r1(pi- xi(t))+ r2(pg- xi(t))] (3)
The second term in Equation (3) can be viewed as an acceleration term, which de-
pends on the distances between the current position xi, the the personal best pi, and the
global best pg. The acceleration factor α is defined as follows:
α = α0 + t/Nt t=1, 2, … , Nt (4)
where Nt denotes the number of iterations, t represents the current generation, and the
suggested range for α0 is [0.5, 1].
As can be seen from Equation (4), the acceleration term will increase as the num-
ber of iterations increases, which will enhance the global search ability at the end of
run and help the algorithm to jump out of the local optimum, especially in the case of
multi-modal problems.
Furthermore, instead of using a linearly-decreasing inertia weight, we use a ran-
dom number, which was proved by Zhang et al. [8] to improve the performance of the
PSO in some benchmark functions. Hence, in this study, we change the inertia weight
at every generation via the following formula:
w=w0 +r(1- w0) ; (5)
where w0 ∈ [0, 1] is a positive constant, and r is a random number uniformly distrib-
uted in [0, 1]. The suggested range for w0 is [0, 0.5], which makes the weight w ran-
domly varying between w0 and 1.
In order to evaluate the performance of individual particles, an appropriate evalua-
tion function (or fitness function) should be defined. Instead of using a ‘dominance-
based evaluation’, we simply use a weighted aggregation approach to construct the
evaluation function F for multi-objective optimisation:
m m
F = ∑ wi f i ; ∑w i =1 (6)
i =1 i =1

where m is the number of objectives, i=1, 2, …, m.

To approximate the Pareto front instead of a certain Pareto solution, the weights wi
for each objective are changed systematically and normalised as follows:
m
wi = λi ∑λ i
; λι=U(0,1) (7)
i =1

The function U(0,1) generates a uniformly distributed random number within the
interval [0,1]. In this way, we can obtain a uniformly distributed random weight
combination, which is generated at every iteration. The idea here is to use dynamic
weights instead of fixed weights to obtain the Pareto solutions. This dynamically
weighted aggregation approach was introduced for the selection of the best pi and pg.
We will show that this approach works very well with both multi-modal test problems
and our industry-related problem.
 Adaptive Weighted Particle Swarm Optimisation for Multi-objective Optimal Design 765

Finally, in order to strengthen the convergence properties of the multi-objective op-

timisation, the ‘non-dominated sorting’ technique, which was proposed and improved
by Deb [9]-[10] and then introduced into the PSO based algorithm by Li [11], has
been also used in our AWPSO algorithm. In the light of above considerations, the
proposed algorithm can be summarised as follows:
1. Initialisation.
Set population number N and iteration number Nt. Initialise the position xi and ve-
locity vi of the particles with random numbers within the pre-defined decision vari-
able range. Vmax is set to the upper bound of the decision variables. Set personal
best position pi=xi, iteration counter t=0.
2. Evaluation.
t=t+1. Evaluate each particle in the current population using equation (6). If
Fi(t)<Fi(t-1), then pi=xi. Find Fmin=min{Fi}, and corresponding position Xmin.
Select global best pg = Xmin.
3. New particles generation.
Calculate the new velocity NVi and new position NXi based on the current xi (i=1,
2, …, N), using equations (3) and (2), and the objective function values for all the
new particles. Combine all xi and NXi (2N particles) together and store them in a
temporary list tempList.
4. Non-dominated Sorting.
Identify non-dominated solutions in tempList and store them in a matrix PFront
(Pareto front). Set front number k=1.
a) Remove the non-dominated particles from tempList.
b) k=k+1. Identify non-dominated solutions in the remaining tempList and store
them in a matrix Frontk (front k).
c) Repeat b) and c) until all 2N particles are ranked into different fronts.
5. Select particles for next iteration.
If PFront size >N, then randomly select N particles from PFront and store them as
NextX. Otherwise, store PFront as NextX then randomly select particles in next
front (Frontk) and add them to NextX until NextX size=N.
6. Calculate objective functions values for all particles in NextX and set the NextX as
the current positions x for the next iteration.
7. If all vi<0.1Vmax, execute the following steps, otherwise go to 8).
a) Randomly select 20% particles in current population and randomly change their
positions by 10% of the Vmax. Store them in Xtemp.
b) Evaluate the Xtemp and find the particles which dominate any particles in the
current Pareto front. Use these dominating particles to replace the corresponding
particles in the current x.
c) Repeat a) and b) K times (K=1~10), make sure the number of x dose not exceed
N.
8. If t < Nt, go to 2).
9. Store the non-dominated solutions from the final population and calculate the per-
formance metric values (see equations (8) and (9) in the next section).
766 Mahdi Mahfouf, Min-You Chen, and Derek Arthur Linkens

3 The Multi-objective Optimisation Problems

In order to demonstrate the effectiveness of the proposed AWPSO algorithm, we used
a set of commonly recognised test problems (ZDT1~ZDT4 functions) [12]. The func-
tion ZDT1 has a convex Pareto front while ZDT2 has a concave Pareto front. Dis-
creteness in the Pareto front for ZDT3 causes difficulties in finding a diverse set of
solutions. ZDT4 is a multi-modal problem, the multiple local Pareto fronts cause
difficulties for many algorithms to converge to the true Pareto-optimal front. For all
test functions, the proposed AWPSO algorithm performed very well and converged to
the Pareto-optimal with a high accuracy while maintaining a good diversity among
the Pareto solutions.
For ZDT1~ZDT3, we set w0=0.15, α0 = 0.5, the population size N=100, the num-
ber of iterations Nt=100. The Pareto solutions found by AWPSO after a single run are
shown in Fig1.(a) to (c) respectively. For the ZDT4 function, the parameters setting
are w0=0.15, α0 = 0.5, N=100, Nt=300. Figure 1 (d) shows the Pareto solutions found
by the AWPSO algorithm. It can be seen that the algorithm possesses very good con-
vergence properties while maintaining a good diversity among the Pareto solutions.
To compare the performance of the AWPSO to other recently developed algorithms,
such as NSGA II, SPEA and NSPSO, two performance metrics, namely the Genera-
tional Distance (GD) and the Spread ∆, which were described in [12], were used. GD
measures the closeness of the obtained Pareto solution set Q from a known set of the
Pareto-optimal set P*, which is defined as follows:
(∑i =1 d im )1 / m ;
|Q|

GD = (8)
|Q|
For a two-objective problem (m=2), di is the Euclidean distance between the solution
i ∈ Q and the nearest member of P*. A set of | P*| =500 uniformly distributed Pareto-
optimal solutions is used to calculate the closeness metric GD.
The Spread ∆ measures the diversity of the solutions along the Pareto front in the
final population and is defined as follows:
∑ d me + ∑i =1 | d i − d |
M |Q|

∆= m =1
(9)
∑
M
m =1
d me + | Q | d
where di is distance between the neighbouring solutions in the Pareto solution set Q.
d is the mean value of all di. d me is the distance between the extreme solutions of P*
and Q along the mth objective. It is worth noting that for an ideal distribution of the
solutions (uniform distribution), ∆=0.
In order to establish repeatability, the AWPSO algorithm was run 10 times inde-
pendently. The average performance metric values and the corresponding variance σ2
are summarised in Tables 1 and 2 respectively. In these Tables, the compared results
for SPEA, NSGA-II and NSPSO were obtained from [12] and [11] respectively. It can
be seen that the proposed algorithm performed very well as far as convergence and
diversity are concerned. From Table 1 we can see that AWPSO has achieved a better
convergence. Table 2 shows that NSGA II has achieved better ∆ values overall. How-
ever, from the Pareto fronts as shown in Fig. 1, AWPSO still has a good coverage of
P*.
 Adaptive Weighted Particle Swarm Optimisation for Multi-objective Optimal Design 767

(a) (b)

(c) (d)
Fig. 1. Pareto solutions of AWPSO on ZDT1~ZDT4.

Table 1. Mean and variance values of the convergence measure GD.

ZDT1 ZDT2 ZDT3 ZDT4
Algor-
ithm GD σ2 GD σ2 GD σ2 GD σ2

SPEA 1.25e-3 0 3.04e-3 2.00e-5 4.42e-2 1.90e-5 9.514 11.321

NSGA II 8.94e-4 0 8.24e-4 0 4.34e-2 4.20e-5 2.92e-2 4.67e-2
NSPSO 7.53e-4 4.18e-5 8.05e-4 3.05e-5 3.4e-3 2.54e-4 7.82e-4 6.91e-5
AWPSO 1.01e-4 2.61e-9 1.210e-4 1.4e-9 5.206e-4 2.85e-9 5.34e-4 3.92e-8

Table 2. Mean and variance values of the diversity measure ∆.

ZDT1 ZDT2 ZDT3 ZDT4
Algorithm 2 2 2
∆ σ∆ ∆ σ∆ ∆ σ∆ ∆ σ∆2
SPEA 0.730 9.07e-3 0.678 4.48e-3 0.666 6.66e-4 0.732 1.13e-2
NSGA II 0.463 4.16e-2 0.435 2.46e-2 0.576 5.08e-3 0.655 1.98e-1
NSPSO 0.767 3.00e-2 0.758 2.77e-2 0.869 5.81e-2 0.768 3.57e-2
AWPSO 0.759 1.71e-3 0.758 4.92e-3 0.891 1.10e-2 0.680 3.00e-3
768 Mahdi Mahfouf, Min-You Chen, and Derek Arthur Linkens

4 Multi-objective Optimisation for Alloy Design

Having established the effectiveness of the algorithm, we decided to apply it to the
optimal design for heat-treated alloy steels. In the steel industry, finding an optimal
design for alloy steels is a multi-objective optimisation problem [13]. Usually, some
objectives may conflict with each other, such as the tensile strength and ductility. In
this section, details relating to the optimisation of tensile strength (TS) and Reduction
of Area (ROA) using the AWPSO algorithm are presented and discussed. All parame-
ters settings in the AWPSO are kept the same as in the previous cases ZDT1~ZDT3.
The decision vector consists of the weight percentages for the chemical composites,
namely: Carbon (C), Silica (Si), Manganese (Mn), Chromium (Cr), Molybdenum
(Mo), Vanadium (V), as well as the Quenching (QT) and Tempering (Temp) Tem-
peratures. As aforementioned, all optimisation experiments are based on the neural-
fuzzy property prediction models [14]. Based on 3760 industrial data, 50% for train-
ing and 50% for model testing, the fuzzy models for TS and ROA prediction were
developed. The prediction results with standard deviation (SD) of the prediction error
are shown in Figs. 2 and 3 respectively. All alloy design optimisation experiments are
based on these fuzzy models.

Fig. 2. Fuzzy model prediction for TS. Fig. 3. Fuzzy model prediction for ROA.

4.1 The Two-Objective Optimisation Problem

In order to achieve the pre-defined tensile strength (TS) and the corresponding 95%
confidence error band EBt for model predictions, such variables were selected as the
objectives. It is worth noting that the error band, which depends upon the model accu-
racy and training data density, provides an accurate guide with respect to the model
prediction error. The smaller the error band, the more reliable the corresponding
model prediction. Hence, the objectives are defined as follows:
 TS − TSt if TS − TSt < 0.1TST
Minimise J1 = 
1000 TS − TSt otherwise (10)
Minimise J 2 = EBT
The first objective function J1 indicates that the ideal solutions should be close to the
target TS value TSt and the acceptable variation is 10% of TSt. The penalty will be
 Adaptive Weighted Particle Swarm Optimisation for Multi-objective Optimal Design 769

given to the solutions that are greater than the 10% variation range of the target. Fig. 4
illustrates the TS values against the error band corresponding to the obtained Pareto
solutions with the target value TSt=868 (N/mm2). It indicates that based on the Pareto
solutions, the obtained TS values are close to the target. Table 3 shows the average
values of the Pareto solutions in all five independent runs. It can be seen that the pro-
duced solutions are very consistent and always converged to a specific area that
minimised the above objective functions.

Fig. 4. Obtained Pareto solutions in the objective space.

Table 3. Mean values of the Pareto solutions across five runs.

C Si Mn Cr Mo Ni V QT Temp TS EB
Run1 0.388 0.251 0.761 1.219 0.304 0.642 0.009 867 650 870 5.06
Run2 0.379 0.252 0.780 1.027 0.271 0.688 0.009 856 626 872 8.69
Run3 0.391 0.256 0.765 1.247 0.279 0.689 0.003 859 642 873 4.81
Run4 0.387 0.250 0.741 1.194 0.308 0.709 0.013 859 653 873 8.01
Run5 0.392 0.256 0.718 0.938 0.261 0.882 0.006 858 619 874 5.76

4.2 The Four-Objective Optimisation Problem

This experiment aims at finding the optimal chemical compositions and heat-
treatment process parameters to obtain the required TS and ROA, the latter reflecting
the ductility of steels. Again, the model prediction error bands EBt and EBr for TS and
ROA respectively are included in the objective functions. The target values for TS
and ROA are set to 868 (N/mm2) and 60%, respectively. Among the four objectives,
J1 and J2 are defined similarly to the previous section, while J3 and J4 relating to ROA
and the corresponding error band are defined as follows:
Minimise: J3=|ROA-60|, J4=EBr (11)
Figure 5 shows the Pareto-solutions produced by the AWPSO algorithm in the
objective space. Again, it is seen that the algorithm converged to the region close to
the pre-defined TS and ROA target values and also provided different solutions which
meet the mechanical property requirements of the alloy steels.
770 Mahdi Mahfouf, Min-You Chen, and Derek Arthur Linkens

Fig. 5. Pareto-solutions in the four-objective space.

In this experiment, we also keep the number of particles to N=100. This indicates
that the algorithm is not sensitive to the population size even in a high-dimensional
objective space.

5 Discussions and Conclusions

In this paper we proposed an adaptive weighted PSO algorithm (AWPSO) and ap-
plied it to a series of benchmark functions and to the optimal design of heat-treated
alloy steels based on data-driven neural-fuzzy predictive models. The main aim of the
research was to determine the optimal heat treatment regime and the required weight
percentages for the chemical composites to obtain the desired mechanical properties
of steels. Using the adaptive weights w0 and α in the AWPSO we overcame the prob-
lem commonly found in the standard PSO algorithm which related to its limitation for
effective local search in the early stages of the run coupled with its limitation for
effective global search during the late stages of the run. The introduction of the adap-
tive weighted aggregation evaluation function improved the diversity of the Pareto
solutions while keeping the algorithm simple. Compared to other recently developed
algorithms, such as NSGA II, SPEA and NSPSO, the proposed algorithm can achieve
better convergence while keeping a good diversity (we used the well-known functions
ZDT1~ZDT4 for benchmark testing). The proposed algorithm has also been success-
fully applied to multi-objective optimal design of alloy steels. The results have shown
that the algorithm can locate the constrained optimal design with very good accuracy,
and other simulation experiments (not reported here) have also established the superi-
ority of the algorithm when compared to the SPEA2 algorithm. The simulation results
also indicate that the algorithm produced very consistent solutions and can be effec-
tively applied to other industrial optimisation problems.

Acknowledgement
The authors would like to thank the UK Engineering and Physical Sciences Research
Council (EPSRC) for their financial support via the IMMPETUS Phase II Programme
Award, under Grant GR/R70514/01.
 Adaptive Weighted Particle Swarm Optimisation for Multi-objective Optimal Design 771

References
1. J. Kennedy and R. Eberhart, Particle Swarm Optimization, Neural Networks, Perth; Aus-
tralia, p. 1942-1948, 1995
2. J. Kennedy, The Particle Swarm: Social adaptation of knowledge, Proc. 1997 Int. Conf.
Evolutionary Computation, Indianapolis, IN, 1997, P. 303-308.
3. J. Kennedy, Stereotyping: Improving Particle Swarm Performance with Cluster Analysis,
Evolutionary Computation, La Jolla, CA, p. 1507-1512, 2000.
4. P.J. Angeline, Evolutionary Optimization Versus Particle Swarm Optimization: Philoso-
phy and Performance Differences, LNCS, p. 601-610, 1998
5. R. Eberhart and Y. Shi, Comparison between Genetic Algorithms and Particle Swarm Op-
timisation, Lecture Notes in Computer Science, p. 611-618, 1998
6. Y. Shi and R. Eberhart, Empirical Study of Particle Swarm Optimization, Evolutionary
Computation, Washington DC, p.1945-1950, 1999
7. Y. Shi and R. Eberhart, Fuzzy Adaptive Particle Swarm Optimization, Congress on
Evolutionary Computation, Seoul, Korea, p. 101-106, 2001.
8. L.H. Zhang and S. Hu, A New Approach to Improve Particle Swarm Optimization, Lec-
ture Notes in Computer Science, vol. 2723, p. 134-139, 2003.
9. K. Deb, S. Agrawal, A. Pratap and T. Meyarivan, A Fast Elitist Non-Dominated Sorting
Genetic Algorithm for Multi-Objective Optimization: NSGA-II, Proceedings of Parallel
Problem Solving from Nature-PPSN VI, Springer, p.849-858, 2000.
10. K. Deb and T. Goel, Controlled Elitist Non-Dominated Sorting Genetic Algorithms for
Better Convergence, Lecture Notes in Computer Science, vol. 1993, p. 67-81, 2001.
11. X. Li, A Non-dominated Sorting Particle Swarm Optimizer for Muli-objective Optimiza-
tion, Lecture Notes in Computer Science, vol. 2723, p. 37-48, 2003.
12. K. Deb, Multi-objective optimization using evolutionary algorithms, New York; Chiches-
ter: Wiley, 2001.
13. M. Mahfouf, M. Jamei, and D.A. Linkens, Optimal Design of Metals using Fuzzy Speci-
fied Multi-Objective Functions, IFAC Fuzzy-GA 2004.
14. M-Y Chen and D.A. Linkens, A Systematic Neuro-fuzzy Modeling Framework with Ap-
plication to Material Property Prediction, IEEE Transactions on Systems, Man and Cyber-
netics, Part B, 31(5), p. 781-790, 2001.