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

net/publication/220704568

Fireworks Algorithm for Optimization

Conference Paper · June 2010

DOI: 10.1007/978-3-642-13495-1_44 · Source: DBLP

CITATIONS READS
987 10,988

2 authors, including:

Ying Tan
Peking University
333 PUBLICATIONS 7,980 CITATIONS

SEE PROFILE

All content following this page was uploaded by Ying Tan on 16 May 2014.

The user has requested enhancement of the downloaded file.

 Fireworks Algorithm for Optimization

Ying Tan and Yuanchun Zhu

Key Laboratory of Machine Perception (MOE), Peking University

Department of Machine Intelligence,
School of Electronics Engineering and Computer Science, Peking University,
Beijing, 100871, China
{ytan,ychzhu}@pku.edu.cn

Abstract. Inspired by observing fireworks explosion, a novel swarm in-

telligence algorithm, called Fireworks Algorithm (FA), is proposed for
global optimization of complex functions. In the proposed FA, two types
of explosion (search) processes are employed, and the mechanisms for
keeping diversity of sparks are also well designed. In order to demon-
strate the validation of the FA, a number of experiments were conducted
on nine benchmark test functions to compare the FA with two vari-
ants of particle swarm optimization (PSO) algorithms, namely Standard
PSO and Clonal PSO. It turns out from the results that the proposed FA
clearly outperforms the two variants of the PSOs in both convergence
speed and global solution accuracy.

Keywords: natural computing, swarm intelligence, fireworks algorithm,

particle swarm optimization, function optimization.

1 Introduction
In recent years, Swarm Intelligence (SI) has become popular among researchers
working on optimization problems all over the world [1,2]. SI algorithms, e.g. Par-
ticle Swarm Optimization (PSO) [3], Ant System [4], Clonal Selection Algorithm
[5], and Swarm Robots [6], etc., have advantages in solving many optimization
problems. Among all the SI algorithms, PSO is one of the most popular algorithm
for searching optimal locations in a D-dimensional space. In 1995, Kennedy and
Eberhart proposed PSO as a powerful global optimization algorithm inspired by
the behavior of bird blocks [3]. Since then, the PSO has attracted the attentions
of researchers around the globe, and a number of variants of PSO have been con-
tinually proposed [7, 8].
Like PSO, most of swarm intelligence algorithms are inspired by some intelli-
gent colony behaviors in nature. In this paper, inspired by the emergent swarm
behavior of fireworks, a novel swarm intelligence algorithm called Fireworks Al-
gorithm (FA) is proposed for function optimization. The FA is presented and
implemented by simulating the explosion process of fireworks. In the FA, two
explosion (search) processes are employed and mechanisms for keeping diversity
of sparks are also well designed. To validate the performance of the proposed
FA, comparison experiments were conducted on nine benchmark test functions

Y. Tan, Y. Shi, and K.C. Tan (Eds.): ICSI 2010, Part I, LNCS 6145, pp. 355–364, 2010.

c Springer-Verlag Berlin Heidelberg 2010
356 Y. Tan and Y.C. Zhu

among the FA, the Standard PSO (SPSO), and the Clonal PSO (CPSO) [8].
It is shown that the FA clearly outperforms the SPSO and the CPSO in both
optimization accuracy and convergence speed.
The remainder of this paper is organized as follows. Section 2 describes the
framework of the FA and introduces two types of search processes and mecha-
nisms for keeping diversity. In Section 3, experimental results are presented to
validate the performance of the FA. Section 4 concludes the paper.

2 Fireworks Algorithm

2.1 FA Framework

When a firework is set off, a shower of sparks will fill the local space around the
firework. In our opinion, the explosion process of a firework can be viewed as
a search in the local space around a specific point where the firework is set off
through the sparks generated in the explosion. When we are asked to find a point
xj satisfying f (xj ) = y, we can continually set off ‘fireworks’ in potential space
until one ‘spark’ targets or is fairly near the point xj . Mimicking the process of
setting off fireworks, a rough framework of the FA is depicted in Fig. 1.
In the FA, for each generation of explosion, we first select n locations, where
n fireworks are set off. Then after explosion, the locations of sparks are obtained
and evaluated. When the optimal location is found, the algorithm stops. Oth-
erwise, n other locations are selected from the current sparks and fireworks for
the next generation of explosion.
From Fig. 1, it can be seen that the success of the FA lies in a good design
of the explosion process and a proper method for selecting locations, which are
respectively elaborated in subsection 2.2 and subsection 2.3.

Select n initial locations

Set off n fireworks at n locations

Obtain the locations of sparks

Evaluate the quality of the locations

Optimal location
found Yes

No
Select n locations End

Fig. 1. Framework of fireworks algorithm

 Fireworks Algorithm for Optimization 357

2.2 Design of Fireworks Explosion

Through observing fireworks display, we have found two specific behavior of

fireworks explosion. When fireworks are well manufactured, numerous sparks are
generated, and the sparks centralize the explosion center. In this case, we enjoy
the spectacular display of the fireworks. However, for a bad firework explosion,
quite few sparks are generated, and the sparks scatter in the space.
The two manners are depicted in Fig. 2. From the standpoint of a search
algorithm, a good firework denotes that the firework locates in a promising area
which may be close to the optimal location. Thus, it is proper to utilize more
sparks to search the local area around the firework. In the contrast, a bad firework
means the optimal location may be far from where the firework locates. Then,
the search radius should be larger. In the FA, more sparks are generated and
the explosion amplitude is smaller for a good firework, compared to a bad one.

(a) Good explosion (b) Bad explosion

Fig. 2. Two types of fireworks explosion

Number of Sparks. Suppose the FA is designed for the general optimization

problem:

Minimize f (x) ∈ R, xmin  x  xmax , (1)

where x = x1 , x2 , . . . , xd denotes a location in the potential space, f (x) is an
objective function, and xmin and xmax denote the bounds of the potential space.
Then the number of sparks generated by each firework xi is defined as follows.

ymax − f (xi ) + ξ
si = m ·  n , (2)
i=1 (ymax − f (xi )) + ξ
where m is a parameter controlling the total number of sparks generated by
the n fireworks, ymax = max(f (xi )) (i = 1, 2, . . . , n) is the maximum (worst)
value of the objective function among the n fireworks, and ξ, which denotes the
smallest constant in the computer, is utilized to avoid zero-division-error.
To avoid overwhelming effects of splendid fireworks, bounds are defined for
si , which is shown in Eq. 3.
358 Y. Tan and Y.C. Zhu

⎧
⎨ round(a · m)
⎪ if si < am
ŝi = round(b · m) if si > bm , a < b < 1 , (3)
⎪
⎩
round(si ) otherwise
where a and b are const parameters.
Amplitude of Explosion. In contrast to the design of sparks number, the am-
plitude of a good firework explosion is smaller than that of a bad one. Amplitude
of explosion for each firework is defined as follows.

f (xi ) − ymin + ξ
Ai = Â · n , (4)
i=1 (f (xi ) − ymin ) + ξ

where  denotes the maximum explosion amplitude,and ymin = min(f (xi )) (i =

1, 2, . . . , n) is the minimum (best) value of the objective function among the n
fireworks.
Generating Sparks. In explosion, sparks may undergo the effects of explosion
from random z directions (dimensions). In the FA, we obtain the number of the
affected directions randomly as follows.
z = round(d · rand(0, 1)) , (5)
where d is the dimensionality of the location x, and rand(0, 1) is an uniform
distribution over [0,1].
The location of a spark of the firework xi is obtained using Algorithm 1.
Mimicking the explosion process, a spark’s location x̃j is first generated. Then
if the obtained location is found to fall out of the potential space, it is mapped
to the potential space according to the algorithm.

Algorithm 1. Obtain the location of a spark

Initialize the location of the spark: x˜j = xi ;
z=round(d · rand(0, 1));
Randomly select z dimensions of x˜j ;
Calculate the displacement: h = Ai · rand(−1, 1);
for each dimension x̃jk ∈ {pre-selected z dimensions of x˜j } do
x̃jk = x̃jk +h;
if x̃jk < xmink or x̃jk > xmax
k then
map x̃k to the potential space: x̃jk = xmin
j j max
k + | x̃k | %(xk − xmin
k );
end if
end for

To keep the diversity of sparks, we design another way of generating sparks —

Gaussian explosion, which is show in Algorithm 2. A function Gaussian(1, 1),
which denotes a Gaussian distribution with mean 1 and standard deviation 1, is
utilized to define the coefficient of the explosion. In our experiments, m̂ sparks
of this type are generated in each explosion generation.
 Fireworks Algorithm for Optimization 359

Algorithm 2. Obtain the location of a specific spark

Initialize the location of the spark: x̂j = xi ;
z=round(d · rand(0, 1));
Randomly select z dimensions of x̂j ;
Calculate the coefficient of Gaussian explosion: g = Gaussian(1, 1);
for each dimension x̂jk ∈ {pre-selected z dimensions of x̂j } do
x̂jk = x̂jk · g;
if x̂jk < xmink or x̂jk > xmax
k then
map x̂k to the potential space: x̂jk = xmin
j j max
k + | x̂k | %(xk − xmin
k );
end if
end for

2.3 Selection of Locations

At the beginning of each explosion generation, n locations should be selected for

the fireworks explosion. In the FA, the current best location x∗ , upon which the
objective function f (x∗ ) is optimal among current locations, is always kept for
the next explosion generation. After that, n − 1 locations are selected based on
their distance to other locations so as to keep diversity of sparks. The general
distance between a location xi and other locations is defined as follows.
 
R(xi ) = d(xi , xj ) =  xi − xj  , (6)
j∈K j∈K

where K is the set of all current locations of both fireworks and sparks.
Then the selection probability of a location xi is defined as follows.

R(xi )
p(xi ) =  . (7)
j∈K R(xj )

When calculating the distance, any distance measure can be utilized including
Manhattan distance, Euclidean distance, Angle-based distance, and so on [9].
When d(xi , xj ) is defined as | f (xi ) − f (xj ) |, the probability is equivalent to
the definition of the immune density based probability in Ref. [10].

2.4 Summary

Algorithm 3 summarizes the framework of the FA. During each explosion genera-
tion, two types of sparks are generated respectively according to Algorithm 1 and
Algorithm 2. For the first type, the number of sparks and explosion amplitude
depend on the quality of the corresponding firework (f (xi )). In the contrast, the
second type is generated using a Gaussian explosion process, which conducts
search in a local Gaussian space around a firework. After obtaining the locations
of the two types of sparks, n locations are selected for the next explosion gener-
ation. In the FA, approximate n + m + m̂ function evaluations are done in each
generation. Suppose the optimum of a function can be found in T generations,
then we can deduce that the complexity of the FA is O(T ∗ (n + m + m̂)).
360 Y. Tan and Y.C. Zhu

Algorithm 3. Framework of the FA

Randomly select n locations for fireworks;
while stop criteria=false do
Set off n fireworks respectively at the n locations:
for each firework xi do
Calculate the number of sparks that the firework yields: ŝi , according to Eq. 3;
Obtain locations of ŝi sparks of the firework xi using Algorithm 1;
end for
for k=1:m̂ do
Randomly select a firework xj ;
Generate a specific spark for the firework using Algorithm 2;
end for
Select the best location and keep it for next explosion generation;
Randomly select n − 1 locations from the two types of sparks and the current
fireworks according to the probability given in Eq. 7;
end while

3 Experiments
3.1 Benchmark Functions
To investigate the performance of the proposed FA, we conducted experiments
on nine benchmark functions. The feasible bounds for all functions are set as
[−100, 100]D . The expression of the functions, initialization intervals and di-
mensionalities are listed in Table 1.

Table 1. Nine benchmark functions utilized in our experiments

Function Expression Initialization D


Sphere F1 = D i=1 x i
2
[30, 50]D 30
 D−1
Rosenbrock F2 = i=1 (100(x i+1 − x2i )2 + (xi − 1)2 ) [30, 50]D 30

Rastrigrin F3 = D 2
i=1 (x i − 10 cos(2πxi ) + 10) [30, 50]D 30
D x 2i 
Griewank F4 = 1 + i=1 4000 − D x
√i
i=1 cos( i ) [30, 50]D 30
D i−1
4 D−1
Ellipse F5 = i=1 10 x2i [15, 30]D 30

Cigar F6 = x21 + D 10 4 2
xi [15, 30]D 30

i=2
Tablet F7 = 104 x21 + D i=2 xi
2
[15, 30]D 30
D
Schwefel F8 = i=1 ((x1 − xi ) + (xi − 1)2 )
2 2
[15, 30]D 30
1
D 2
F9 = 20 + e − 20exp −0.2 D i=1 x i
Ackley [15, 30]D 30
1
 D 2
−exp D i=1 cos(2πx i )

3.2 Comparison Experiments among the FA, the CPSO and the
SPSO
In this section, we compare the performance of the FA with the CPSO and the
SPSO in terms of both convergence speed and optimization accuracy.
 Fireworks Algorithm for Optimization 361

Table 2. Statistical mean and standard deviation of solutions found by the FA, the
CPSO and the SPSO on nine benchmark functions over 20 independent runs

Function FA’s mean CPSO’s mean SPSO’s mean

Function
evluations (StD) (StD) (StD)
0.000000 0.000000 1.909960
Sphere 500000
(0.000000) (0.000000) (2.594634)
9.569493 33.403191 410.522552
Rosenbrock 600000
(12.12829) (42.513450) (529.389139)
0.000000 0.053042 167.256119
Rastrigrin 500000
(0.000000) (0.370687) (42.912873)
0.000000 0.632403 2.177754
Griewank 200000
(0.000000) (0.327648) (0.294225)
0.000000 0.000000 53.718807
Ellipse 500000
(0.000000) (0.000000) (68.480173)
0.000000 0.000000 0.002492
Cigar 600000
(0.000000) (0.000000) (0.005194)
0.000000 0.000000 1.462832
Tablet 500000
(0.000000) (0.000000) (1.157021)
0.000000 0.095099 0.335996
Schwefel 600000
(0.000000) (0.376619) (0.775270)
0.000000 1.683649 12.365417
Ackley 200000
(0.000000) (1.317866) (1.265322)

The parameters of both the CPSO and the SPSO are set as those in Ref. [8].
For the FA, the parameters were selected by some preliminary experiments. We
found that the FA worked quite well at the setting: n = 5, m = 50, a = 0.04,
b = 0.8, Â = 40, and m̂ = 5, which is applied in all the comparison experiments.
Table 2 depicts the optimization accuracy of the three algorithms on nine
benchmark functions, which are averaged over 20 independent runs. It can be
seen that the proposed FA clearly outperforms both the CPSO and SPSO on all
the functions. In addition, the FA can find optimal solutions on most benchmark
functions in less than 10000 function evaluations, as shown in Table 3. However,
the optimization accuracy of the CPSO and the SPSO is unacceptable within
10000 function evaluations.
Besides optimization accuracy, convergence speed is quite essential to an opti-
mizer. To validate the convergence speed of the FA, we conducted more thorough
experiments. Fig. 3 depicts the convergence curves of the FA, the CPSO and the
SPSO on eight benchmark functions averaged over 20 independent runs. From
these results, we can arrive at a conclusion that the proposed FA has a much
faster speed than the CPSO and the SPSO. From Table 3, we can find that the
FA can find excellent solutions with only 10000 times of function evaluations.
This also reflects the fast convergence speed of the proposed FA.
362 Y. Tan and Y.C. Zhu

4 9
x 10 x 10
5
FWA 8 FWA
4 CPSO CPSO
SPSO 6 SPSO

Best fitness
Best fitness

3
4
2
2
1

0 0

0 2000 4000 6000 8000 10000 0 0.5 1 1.5 2

Function evaluations Function evaluations 4
x 10
(a) Sphere (b) Rosenbrock
4
x 10
5
FWA FWA
4 CPSO CPSO
SPSO 10 SPSO
Best fitness
Best fitness

2 5

0 0

0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000
Function evaluations Function evaluations
(c) Rastrigin (d) Griewank
6 7
x 10 x 10
20
FWA 15
CPSO
15 SPSO
Best fitness

Best fitness

10
10
FWA
5 CPSO
5
SPSO

0 0

0 2000 4000 6000 8000 0 2000 4000 6000 8000

Function evaluations Function evaluations
(e) Ellipse (f) Cigar
5 6
x 10 x 10
3
FWA 8
CPSO
2 SPSO 6
Best fitness
Best fitness

4
1 FWA
2 CPSO
SPSO
0 0

0 2000 4000 6000 0 2 4 6 8 10

Function evaluations Function evaluations 4
x 10
(g) Tablet (h) Schwefel

Fig. 3. Convergence curves of the FA, the CPSO and the SPSO on eight benchmark
functions. The function fitness are averaged over 20 independent runs.
 Fireworks Algorithm for Optimization 363

Table 3. Statistical mean and standard deviation of solutions found by the FA, the
CPSO and the SPSO on nine benchmark functions over 20 independent runs of 10000
function evaluations

FA’s mean CPSO’s mean SPSO’s mean

Function
(StD) (StD) (StD)
0.000000 11857.425781 24919.099609
Sphere
(0.000000) (3305.973067) (3383.241523)
19.38330 2750997504.000000 5571942400.000000
Rosenbrock
(11.94373) (1741747548.420642) (960421617.568024)
0.000000 10940.148438 24013.001953
Rastrigrin
(0.000000) (3663.484331) (4246.961530)
0.000000 3.457273 7.125976
Griewank
(0.000000) (0.911027) (0.965788)
0.000000 2493945.500000 5305106.500000
Ellipse
(0.000000) (1199024.648305) (1117954.409340)
0.000000 122527168.000000 149600864.000000
Cigar
(0.000000) (28596381.089661) (13093322.778560)
0.000000 15595.107422 42547.488281
Tablet
(0.000000) (8086.792234) (8232.221882)
4.353733 8775860.000000 6743699.000000
Schwefel
(1.479332) (1217609.288290) (597770.084232)
0.000000 15.907665 18.423347
Ackley
(0.000000) (1.196082) (0.503372)

3.3 Discussion
As shown in the experiments, the FA has a faster convergence speed and a better
optimization accuracy, compared to the PSO. We consider the success of the FA
lies in the following two aspects.
– In the FA, sparks suffer the power of explosion from z dimensions simul-
taneously, and the z dimensions are randomly selected for each spark x̃i .
Thus, there is a probability that the differences between the firework and
the target location happen to lie in these z dimensions. In this scenario, the
sparks of the firework can move towards the target location from z directions
simultaneously, which endues the FA with a fast convergence speed.
– Two types of sparks are generated to keep the diversity of sparks, and the
specific selection process for locations is a mechanism for keeping diversity.
Therefore, the FA has the capability of avoiding premature convergence.

4 Conclusions
Mimicking the explosion process of fireworks, the so-called FA is proposed and
implemented for function optimization. The experiments among the FA, the
CPSO and the SPSO have shown that the proposed FA has a promising per-
formance. It clearly outperforms the CPSO and the SPSO on nine benchmark
364 Y. Tan and Y.C. Zhu

functions in terms of both optimization accuracy and convergence speed, which

endues the FA with a promising prospect of application and extension. In future
work, we will seek a deep theoretical analysis on the FA and try to apply the
FA to some practical engineering applications. Finally, we intend to discuss the
relationship between the FA and other general-purpose optimization algorithms.

Acknowlegements. This work was supported by National Natural Science

Foundation of China (NSFC), under Grant No. 60875080 and 60673020, and
partially supported by the National High Technology Research and Develop-
ment Program of China (863 Program), with Grant No. 2007AA01Z453. Authors
appreciate Mr. Chao Rui for his facility with the early phase of the experiments.

References
1. Garnier, S., Gautrais, J., Theraulaz, G.: The biological principles of swarm intelli-
gence. Swarm Intelligence 1(1), 3–31 (2007)
2. Das, S., Abraham, A., Konar, A.: Swarm intelligence algorithms in bioinformatics.
Studies in Computational Intelligence 94, 113–147 (2008)
3. Kennedy, J., Eberhart, R.C.: Particle swarm optimization. In: Proceedings of IEEE
International Conference on Neural Networks, vol. 4, pp. 1942–1948 (1995)
4. Dorigo, M., Maniezzo, V., Colorni, A.: Ant system: optimization by a colony of
cooperating agents. IEEE Transactions on Systems, Man, and Cybernetics, Part
B: Cybernetics 26(1), 29–41 (1996)
5. De Castro, L.N., Von Zuben, F.J.: Learning and optimization using the
clonal selection principle. IEEE Transactions on Evolutionary Computation 6(3),
239–251 (2002)
6. Beni, G., Wang, J.: Swarm intelligence in cellular robotic systems. In: Proceedings
of NATO Advanced Workshop on Robots and Biological Systems (1989)
7. Bratton, D., Kennedy, J.: Defining a standard for particle swarm optimization. In:
Proceedings of IEEE Swarm Intelligence Symposium, pp. 120–127 (2007)
8. Tan, Y., Xiao, Z.M.: Clonal particle swarm optimization and its applications.
In: Proceedings of IEEE Congress on Evolutionary Computation, pp. 2303–2309
(2007)
9. Perlibakas, V.: Distance measures for PCA-based face recognition. Pattern Recog-
nition Letters 25(6), 711–724 (2004)
10. Lu, G., Tan, D., Zhao, H.: Improvement on regulating definition of antibody density
of immune algorithm. In: Proceedings of the 9th International Conference on Neural
Information Processing, vol. 5, pp. 2669–2672 (2002)

View publication stats