Applied Mathematics and Computation: Hanning Chen, Yunlong Zhu

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 14

Applied Mathematics and Computation 205 (2008) 47–60

Contents lists available at ScienceDirect

Applied Mathematics and Computation


journal homepage: www.elsevier.com/locate/amc

Optimization based on symbiotic multi-species coevolution


Hanning Chen a,b,*, Yunlong Zhu a
a
Shenyang Institute of Automation, Chinese Academy of Sciences, Faculty Office II, Nanta Street 114#, Dongling District, Shenyang 110016, China
b
Graduate School of the Chinese Academy of Sciences, Yuquan Road (Jia) 19#, Shijingshan District, Beijing 100039, China

a r t i c l e i n f o a b s t r a c t

Keywords: This paper presents a general optimization model gleaned ideas from the coevolution of
Symbiosis symbiotic species in natural ecosystems. Species extinction and speciation events are also
Multi-species coevolution considered in this model to tie it closer to natural evolution, as well as improve the algo-
Species extinction and speciation rithm robustness. This model is instantiated as a novel multi-species optimizer, namely
PSO
PS2O, which extends the dynamics of the canonical PSO algorithm by adding a significant
PS2O
ingredient that takes into account the symbiotic coevolution between species. When tested
against benchmark functions, the PS2O markedly outperforms the canonical PSO algorithm
in terms of accuracy, robustness and convergence speed.
Ó 2008 Elsevier Inc. All rights reserved.

1. Introduction

Nature ecosystems have always been the rich source of mechanisms for designing computational systems to solve diffi-
cult engineering and computer science problems. Cooperation between entities is such a universal mechanism. Nature is full
of examples of cooperation: within species (i.e., homogeneous cooperation, also called social evolution), as in the social for-
aging behaviours of animal herds, bird flocks, insect groups and bacterial colonies; between species (i.e., heterogeneous
cooperation, also called symbiosis), as in the mutualism between human and honeyguide (a kind of bird) – bird guides hu-
man to bee’s nest and human open hive so bird can eat beeswax and dead bees.
In recently years, drew inspiration form the homogeneous cooperation within species, the most successful computational
system – Swarm Intelligence (SI), has been presented. Currently, two best developed SI paradigms can be found in the lit-
erature [1–4], namely Ant Colony Optimization (ACO) and Particle Swarm Optimization (PSO). In the ACO paradigm, based
on trail formation via pheromone deposition/evaporation, artificial ants are employed to cooperatively find good solutions
for discrete optimization problems. PSO gleaned ideas from social behaviour of bird flocking or fish schooling. This paradigm
is primarily concerned with continuous optimization and has already come to be widely used in many areas [5,6,25].
However, like the previous EAs (i.e., Evolutionary Algorithms, which drew inspiration from evolution by nature selection),
SI algorithms suffer from the following drawback: as a population evolves, all individuals suffer premature convergence to
the local optimum in the first generations. This leads to low population diversity and adaptation stagnation in successive
generations. However, such loss of population diversity is not observed in natural systems. Because populations of species
interact with one another in natural ecosystems, these species form biological communities which are large social systems
typically consist of both heterogeneous and homogeneous aspects. The interaction between species and the complexity of
their relationships in these communities exemplify what is meant by the term ‘‘biodiversity”.
In this paper, we adopt not only the social evolution perspective (i.e., homogeneous cooperation) but also the symbiosis
theory (i.e., heterogeneous cooperation between species) in formulating our computer simulation models. We introduced a

* Corresponding author. Address: Shenyang Institute of Automation, Chinese Academy of Sciences, Faculty Office II, Nanta Street 114#, Dongling District,
Shenyang 110016, China.
E-mail address: chenhanning@sia.cn (H. Chen).

0096-3003/$ - see front matter Ó 2008 Elsevier Inc. All rights reserved.
doi:10.1016/j.amc.2008.05.148
48 H. Chen, Y. Zhu / Applied Mathematics and Computation 205 (2008) 47–60

Fig. 1. The cooperation levels of PS2O.

number of N species into this model to represent the ‘‘ecosystem”. Two level of cooperation: species level (interaction be-
tween species) and individual level (interaction within species) are simulated in our model (shown as in Fig. 1.). The dynam-
ics of each species in the ecosystem is manipulated by extending the dynamics of the canonical PSO model. We extend the
control law of the canonical PSO model by adding a significant ingredient, which takes into account the symbiotic coevolu-
tion between species. Since mass extinction [10] can be a natural feature of the ecosystem’s dynamics and sometimes is the
result of the coevolution of species, species extinction and speciation are also simulated in our model. This model is instan-
tiated as a novel multi-species optimizer; since this proposed algorithm contains two hierarchies and is based on the PSO
model, we named it as PS2O.
The description of the biological details of the symbiosis theory and mass extinction theory in this paper were taken from
[7–10]. Some tentative concepts for the evolution genetics of symbiosis is proposed in [11]. A mathematical model of the
origin of symbiosis is provided in [12].
We should note that the notion of species or subswarm has been introduced in some variant version of the PSO algorithm.
A multi-species PSO (MS-PSO) was proposed by Chow and Tsui [13], their species optimizes different function thus realizes
the multi-objective optimization. Other variants [14,15] which use species or subswarm have been proposed to encourage
the emergence of niches in a single population. However, the niches represent competition rather than cooperation. Different
cooperative approaches have also been introduced in [16–18]. Frans and Andries used multiple swarms to optimize different
components of the solution vector [16]. The approach introduced in [17,18] relies on having two or more swarms searching
concurrently for a solution with frequent interacting, while the interaction frequency is arbitrarily predefined. A multi-
swarm cooperative particle swarm optimizer (MCPSO) is described in [24,33]. In MCPSO, the population consists of one mas-
ter swarm and several slave swarms. The slave swarms execute the canonical PSO algorithm or its variants independently to
maintain the diversity of particles, while the particles in the master swarm enhance themselves based on their own knowl-
edge and also the knowledge of the particles in the slave swarms.
Our model for the coevolution of symbiotic species is inherently different from past ones in the following ways: (1) this
model contains a number N of species, and each species possesses a certain number M of individuals; (2) all species are sep-
arated breeding population and concurrently search the problem space to obtain parallelism; (3) the evolution of each species
is handled by an identical force generation equation which simulated cooperation not only within species but also between
species; (4) the symbiotic relationship is inherently set between species, and all species interact one other in each generation.
Cooperation is now conducting both within species and between species over the entire life cycle of the ecosystem; (5) spe-
cies extinction events and speciation events are also simulated in our model to address coevolutionary avalanches [19].
Clearly we model more details of the cooperative coevolution mechanism in nature ecosystems and tie this model closer
to natural evolution. In the next section, the canonical PSO model, symbiosis and mass extinction theory are reviewed.
Description of the proposed model is given in Section 3. Next, experimental settings and experimental results are given
in Section 4. Finally, Section 5 concludes the paper.

2. Modeling

2.1. Symbiosis

Symbiosis, initially defined by Anton de Bary in 1879, is simply the living together of organisms from different species.
Here, we denote symbiosis as relationships that are constant and intimate between dissimilar species. Symbiosis includes
three classify types: mutualism, commensalism and parasitism. In detail, mutualism describes a relationship in which all
organisms involved derive benefits; commensalism result in only one species benefits without apparent benefit or cost to
H. Chen, Y. Zhu / Applied Mathematics and Computation 205 (2008) 47–60 49

the other members of the association; and parasitism is a relationship in which one organism benefits at the cost to the other
members. In a sense, what becomes interesting about symbiosis is not so much that organisms live together, but they coop-
erate. Over the years, symbiosis is used to refer to the special case of mutualism. Ultimately, what is of interest is not mutu-
alism, but the cooperation that enable dissimilar species to intimately associate with each other over evolutionary significant
durations.
Symbiosis is almost ubiquitous in nature. There are practically no plants or animals free of symbionts (organisms in sym-
biotic relationship) living on or in them. Research shows different types of symbiotic interactions. Some involve internal
interactions, like bacteria in human intestines. Some of these interactions have seemed to lead to the evolution of organisms
(e.g., the eukaryote cells, from which all plants and animals are descended have symbiotic origin). Others appear to be purely
behavioural, as in a human–honeyguide mutualism that discussed above. Currently, enlightened evolutionary theory recog-
nizes symbiosis as an integral process, and a fundamental source of innovation in evolution.

2.2. Mass extinction

Species extinction has played a significant role in the history of life on Earth. Of the estimated one to four billion species
have existed on the Earth, while less than 50 million are still alive today and all the others became extinct. There are two
primary colleges of thought about the causes of extinction. The traditional view, still held by most palaeontologists as well
as many in other disciplines, is that extinction is the result of external stresses imposed on the ecosystem by the environ-
ment. At the other end of the scale, an increasing number of biologists and ecologists are supporting the idea that extinction
has biotic causes – that extinction is a natural part of the dynamics of ecosystems and would take place regardless of any
stresses arising from the environment. There is evidence in favour of this viewpoint that extinction can be the result of
the interactions and dependencies between species. For example, in symbiotic mutualism or food web interactions context,
the extinction of a specific species can cause the extinction of others.
There are two best known model of species extinction [10]: the Bak-Sneppen model, which attempts to explain mass
extinction as a result of species interactions; the extinction model of Newman, which models extinction as the result of envi-
ronmental influences on species.

2.3. Canonical PSO model

One of the best developed SI systems is Particle Swarm Optimization (PSO), which is a kind of swarm intelligence inspired
by emergent collective intelligence of social model (e.g., bird flocks and fish shoals). The canonical PSO (CPSO) model evolves
a single population of interacting particles, moving around in the D dimensional problem space searching for the optimum.
Each particle is represented as ~ xi ¼ ðxi1 ; xi2 ; . . . ; xiD Þ and records its previous best position represented as Pi = (Pi1, Pi2,. . ., PiD),
which is also called pbest. The index of the best particle among all the particles in the population is represented by the
symbol g, and pg is called gbest. In the canonical PSO model, each particle is accelerated by two elastic forces, i.e., one attracts
it to pbest and the other to gbest. The magnitude of the force is randomly chosen at each time step. At each generation, the
equation controlling the particles is of the form:
ai ¼ c1 r 1 ðpi  xi Þ þ c2 r2 ðpg  xi Þ; ð1Þ

where c1 and c2 are two learning rates that control, respectively, the proportion of social transmission and individual learning
in the swarm and r1, r2 are two random vectors uniformly distributed in [0, 1].
The velocity of particle i, and its position updated every generation using the equations:
vi ¼ vðvi þ ai Þ xi ¼ xi þ vi ; ð2Þ
where v is known as the constriction coefficient [20].
However, studies showed that the canonical PSO (or original PSO algorithm) has difficulties in controlling the balance be-
tween exploration (global investigation of the search place) and exploitation (the fine search around a local optimum). The
CPSO algorithm performs well in the early iterations (i.e., quickly converging towards an optimum in the first period of iter-
ations), while has problems reaching a near optimal solution in some function optimization problems. Various attempts have
been made to improve the performance of CPSO, which can be found in literatures [26–33].

3. Symbiosis coevolution model for optimization

In this section, we describe our model for the coevolution of symbiotic species and formulate it as an optimization algo-
rithm. We present the outline of our model by making the following assumptions:

(1) Within species, species members cooperate with each other and rely on the presence of other species members for
survival.
(2) Between species, symbiotic partners from distinct species cooperate with each other and all partners gain an advan-
tage to increase their survival ability.
50 H. Chen, Y. Zhu / Applied Mathematics and Computation 205 (2008) 47–60

(3) Cooperation both within and between species are obligate through the whole life cycles of all species.
(4) All species feel the same external environmental stress. Some species will become extinct if the stress is severe
enough.
(5) If one or more species went extinct, they will be replaced with equal number of new ones. Thus the number of species
remains constant.

These assumptions yield a model that can be instantiated as the optimization algorithm, namely PS2O, which present
below:
Here the basic goals is to find the minimum of f ð~ x 2 Rd . We create an ecosystem contains a species set X = {S1, S2,. . .,
xÞ; ~
Sn}, and each species possesses a members set Sn ¼ fxn1 ; xn2 ; . . . ; xnm g i.e., totally n  m individuals coevolve in the ecosystem.
The ith member of the kth species is characterized by the vector ~ xki ¼ ðxki1 ; xki2 ; . . . ; xkid Þ. Suppose f ð~
xki Þ is the fitness of xki and
lower value of the fitness represents the higher ability of survival. Under this presumed external environmental stress, all
individuals in our model coevolve to the states of lower and lower fitness by cooperating each other both within species
and between species. In each generation t, each individual xki behaves as follow:
(1) Social evolution: this process addresses the cooperation between individuals of the same species. Due to the socio-
biological background of the canonical PSO model, xki evolve according to the rules of the canonical PSO algorithm in this
process. Within the kth species, one or more members in the neighbourhood of xki contribute their knowledge to xki and xki
also share its knowledge with its neighbours. Then xki accelerate towards the personal best position and the best position
found by its neighbours:

aki ¼ c1 r1 ðpki  xki Þ þ c2 r 2 ðpkg  xki Þ; ð3Þ

where k is the index of the species that the xki


belongs to, aki
is the acceleration vector of xki , pki
is the personal best position
found so far by xki , pkg is the best position found so far by its neighbours within species k, c1 are the individual learning rates, c2
are the social learning rate; and r 1 ; r 2 2 Rd are two random vectors uniformly distributed in [0, 1].
(2) Symbiotic evolution: this process addresses the cooperation between individuals of distinct species. xki beneficially
interacts with and rewards all its symbiotic partners (individuals of dissimilar species), i.e., each symbiotic partner donates
its knowledge to aid other partners. Then xki accelerate towards its symbiotic partner of the best fitness:

aki ¼ aki þ c3 r 3 ðpls  xki Þ; ð4Þ


d
where l is the index of the species which the best symbiotic partner belongs to, c3 is the ‘‘symbiotic learning rate”, r 3 2 R is a
uniform random sequence in the range [0, 1], and pls is the best position found so far by the symbiotic partners of xki .
(3) If all individuals in the ecosystem cannot find a better position after a (here a is a constant) generations, it means that
all species suffer a severe external environmental stress. Then we randomly choose half species of the ecosystem go extinct
to release this stress for other species to survive. At the same time, random initiate equal number of species in the ecosystem
for new experimentations and adaptations. Let e be the set of the index of species that will go extinct, then the velocity vki of
each individual and its position are updated according to

if ðk 2 eÞ xki ¼ r  ðub  lbÞ þ ub; ð5Þ


else vki ¼ vðvki þ aki Þ; xki ¼ xki þ vki ; ð6Þ
where ub and lb are lower and upper boundaries of the search place, and r 2 Rd is a random vectors uniformly distributed in
[0, 1]. In summary, the equations in the three processes above can be combined as follow:
      
vki ¼ v vki þ c1 r1 pki  xki þ c2 r 2 pkg  xki þ c3 r 3 pls  xki ; ð7Þ


r  ðub  lbÞ þ ub if species k went extinct
xki ¼ : ð8Þ
xki þ vki otherwise

The term c1 r1 ðpki  xki Þ is associated with cognition since it takes into account the individual’s own experiences. The term
c2 r2 ðpkg  xki Þ represents the social interaction within species k. The new term adding into the velocity update equation,
c3 r3 ðpls  xki Þ, takes into account the symbiotic coevolution (i.e., shared memory) between dissimilar species. With an extra
control to the position update equation, the species extinction and speciation are also modelled.
In our model, the cooperation occurred in two levels, i.e., species level (interaction between species) and individual level
(interaction within species). Two hierarchical interaction topologies have been employed in this paper to realize this two-
level cooperative mechanism. In the first topology (namely the 2-level global topology, shown as in Fig. 2), each individual
is influenced by the performance of its own species and all the other species in the ecosystem. In the second topology
(namely the 2-level local topology, shown as in Fig. 3), each individual is influenced only by n closest neighbours from its
own species and other n species from the ecosystem. Here, n = 2. Two variant version of the PS2O algorithm are studied
according to the different interaction topologies, respectively.
The flowchart of the PS2O algorithm is presented in Fig. 4. The pseudocode for the PS2O algorithm is listed in
Table 1.
H. Chen, Y. Zhu / Applied Mathematics and Computation 205 (2008) 47–60 51

Fig. 2. The 2-level global interaction topology.

Fig. 3. The 2-level local interaction topology.

Fig. 4. Flowchart of the PS2O algorithm.


52 H. Chen, Y. Zhu / Applied Mathematics and Computation 205 (2008) 47–60

Table 1
Pseudocode for the PS2O algorithm

Set t: = 0;
INITIALIZE. Randomize positions and velocities of n*m particles in search space. Randomly divide the whole population into n species each possesses m
particles;
WHILE (the termination conditions are not met)
FOR (each species k) IN PARALLEL
FOR (each particle i of species k)
Update the velocity using Eq. (7)
END FOR
IF (all species cannot find a better position after a generations)
Randomly choose a extinct species set e
IF (k 2 e)
Update the position using Eq. (5)
END IF
ELSE
Update the position using Eq. (6)
END FOR IN PARALLEL
Set t: = t + 1;
END WHILE

4. Experiment and result

4.1. Benchmark functions

The set of benchmark functions contains six functions that are commonly used in evolutionary computation literature
[21–23] to show solution quality and convergence rate. The first two functions are unimodal problems and the remaining
four functions are multimodal. The functions are listed below. The dimensions, initialization ranges, global optimum, and
the criterion of each function are listed in Table 2.

1. Sphere function
X
n
f1 ðxÞ ¼ x2i : ð9Þ
i¼1

2. Rosenbrock function
X
n
f2 ðxÞ ¼ 100  ðxiþ1  x2i Þ2 þ ð1  xi Þ2 : ð10Þ
i¼1

3. Ackley
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1
0 !
u n
u1 X 1 Xn
f3 ðxÞ ¼ 20 exp @0:2 t x A  exp
2
cos 2pxi þ 20 þ e: ð11Þ
n i¼1 i n i¼1

4. Rastrigrin function
X
n
f4 ðxÞ ¼ ðx2i  10 cosð2pxi ÞÞ þ 10: ð12Þ
i¼1

5. Griewank function

1 X n Yn
xi
f5 ðxÞ ¼ x2i  cos pffi þ 1: ð13Þ
4000 i¼1 i¼1 i

Table 2
Parameters of the test functions

Function Dimensions Initial range Minimum Criterion


Sphere 30 [100, 100]D 0 0.01
Rosenbrock 30 [30, 30]D 0 100
Ackley 30 [32.768, 32.768]D 0 0.1
Rastrigrin 30 [5.12, 5.12]D 0 100
Griewank 30 [600, 600]D 0 0.1
Weierstrass 30 [0.5, 0.5]D 0 10
H. Chen, Y. Zhu / Applied Mathematics and Computation 205 (2008) 47–60 53

6. Weierstrass function
! !
X
n kX
max
k
kX
max
k
f6 ðxÞ ¼ ½ak cosð2pb ðxi þ 0:5ÞÞ  n ½ak cosð2pb  0:5Þ ; ð14Þ
i¼1 k¼0 k¼0

where a = 0.5, b = 3, kmax = 20.

4.2. Settings

Experiments were conducted with the canonical PSO (CPSO) and three variations of PS2O, namely the PS2O-g (using the 2-
level global interaction topology), the PS2O-l (using the 2-level local interaction topology), and the PS2O-ex ( using the 2-
level local interaction topology with the mass extinction operator).
In order to investigate whether the performance of PS2O is sensitive to the number of species or not, we test PS2O-g and
PS2O-l on the first three benchmark function of 30 dimensions by changing the number of species and fixing the whole pop-
ulation size at 100. Parameters were set to the values c1 = c2 = c3 = 1.367, and v = 0.729. The maximum velocity of each par-
ticle was set to be 20% of the search space. The max generation is set at 2000. The average test results obtained from 25 runs
are listed in Table 3. From the results, we could observe when the species numbers increasing, PS2O gives the better perfor-
mance on both the unimodal functions and the multimodal functions. That is, the PS2O model is sensitive to the number of
species.
In Section 4.3, the population size for all algorithms was set at 100. The maximum velocity for all algorithms was set to be
5% of the search space for unimodal functions and 50% for multimodal functions. Two learning rates c1 and c2 for the CPSO
were both 2.05. For all variants of the PS2O, three learning rates were set to the values c1 = c2 = c3 = 1.367, the number of spe-
pffiffiffiffiffiffiffiffiffiffiffiffiffi
cies was set at 20 (i.e., the population size of each species is 5), and the constant a is given by round ð n  mÞ, where n is the
number of species and m is the population size of each species. The parameters setting for all algorithms are summarized in
Table 4.

4.3. Experiment 1: numerical results and comparison

In Experiment 1, we compare three variant versions of the PS2O algorithm with the canonical PSO algorithm in the func-
tion optimization domain. We extracted three kinds of measures of performance on the six benchmark functions. The exper-
iments run 25 times, respectively for each algorithm on each benchmark function and the max generation is set at 10,000.
The first measure is the best function results after 10,000 iterations. The representative results obtained are presented in
Table 5, including the best, worst, mean and standard deviation of the function values found in 25 runs. Fig. 5 through 10
present the evolution process for all algorithms according to the reported results in Table 5.
The second measure includes two dependent variables. One is the number of iterations required to reach the criterion and
the other is the number of iterations required to reach the global optimum. The second measure is also a measure of speed. In
this case, the first dependent variable indicates that the searcher whether can arrive in the region of the global optimum;
while the second dependent variable indicates that the searcher whether can find the global optimum or not. If the goal

Table 3
Results achieved under different number of species

Funtion PS2O-g PS2O-l


N 5 10 20 5 10 20
f1 8.0344e-071 3.4320e-077 1.0866e-086 4.4128e-033 1.6835e-035 4.3836e-036
f2 0.8974 0.7975 0.6029 19.6735 18.8602 16.1160
f3 2.3997 1.7287 1.1354 7.8620e-015 6.2172e-015 6.2172e-015

Table 4
Parameter setting for all algorithms

Type PS2O-g PS2O-l PS2O-ex CPSO


Population size 100 100 100 100
Number of species 20 20 20 NA
v 0.729 0.729 0.729 0.729
c1 1.367 1.367 1.367 2.05
c2 1.367 1.367 1.367 2.05
c3 1.367 1.367 1.367 NA
a NA NA 10 NA
54 H. Chen, Y. Zhu / Applied Mathematics and Computation 205 (2008) 47–60

Table 5
Performance of all algorithms

Function (dim.30) PS2O-g PS2O-l P2SO-ex CPSO


f1 Best 0 3.0873e-183 6.8504e-118 4.2513e-092
Worst 0 3.6064e-179 1.8712e-051 2.0333e-082
Mean 0 2.5549e-180 9.3561e-053 1.2045e-083
Std 0 0 4.1842e-052 4.5752e-083
f2 Best 1.6416e-028 0.0011 0.0021 0.0360
Worst 3.9866 0.0113 3.9958 15.2582
Mean 0.9967 0.0037 0.4095 7.3954
Std 1.7711 0.0032 1.2255 4.7577
f3 Best 2.6645e-015 2.6645e-015 2.6645e-015 2.6645e-015
Worst 2.4954 6.2172e-015 6.2172e-015 1.3404
Mean 1.1363 3.1974e-015 3.0198e-015 0.0670
Std 0.9273 1.3015e-015 1.0935e-015 0.2997
f4 Best 23.8790 6.9647 10.3188 31.8386
Worst 85.5662 29.8488 30.8437 67.6571
Mean 47.7082 23.2323 21.3603 51.7378
Std 14.6636 5.5564 4.8020 11.2427
f5 Best 0 0 0 0
Worst 0.0320 0.0123 0 0.0467
Mean 0.0095 9.8561e-004 0 0.0136
Std 0.0100 0.0031 0 0.0143
f6 Best 3.6226 0 0 3.1138
Worst 9.0740 0 0 4.6955
Mean 5.6615 0 0 4.0214
Std 2.1868 0 0 0.6619

In bold are the best results.

50

-50

-100
Fitness (log)

-150

-200
PS2O-g
-250
PS2O-l

-300 PS2O-ex
CPSO
-350
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Generations
Fig. 5. f1, Sphere function.

(i.e., reach the criterion or the global optimum) was not reached within the maximum number of 10,000 iterations, the run
was considered unsuccessful. For the successful runs, the average number of iterations to achieve the goals was presented in
Table 6. The unsuccessful runs are shown in the table with the lemniscus.
The third measure also includes two dependent variables. One is the proportion of trials that successfully found the cri-
terion and the other is the proportion of trials that successfully found the global optimum. This measure tells whether the
given algorithm can solve the problems. The results are listed in Table 7.
From all the results for the three measures, we can observe that the PS2O algorithm obtain an obviously remarkable per-
formance. We can see it clearly all PS2O variants converged greatly faster and to significantly better results than the CPSO for
both unimodal and multimodal cases. The PS2O-g is the most fast ones for finding good results within relatively few
generations.
H. Chen, Y. Zhu / Applied Mathematics and Computation 205 (2008) 47–60 55

Table 6
Median number of iterations to criteria/minimum

Median number of iterations to criteria Median number of iterations to minimum


PS2O-g PS2O-l PS2O-ex CPSO PS2O-g PS2O-l PS2O-ex CPSO
f1 113 239 251 429 7356 1 1 1
f2 81 104 111 283 1 1 1 1
f3 115 272 313 638 1 1 1 1
f4 65 332 654 178 1 1 1 1
f5 134 317 380 653 1 1046 2183 1
f6 96 126 819 251 1 1683 6156 1

In bold are the best results.

0
Fitness (log)

-1

-2 PS2O-g

PS2O-l
-3
PS2O-ex
CPSO
-4
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Generations

Fig. 6. f2, Rosenbrock function.

-2

-4
PS2O-g
-6
Fitness (log)

PS2O-l
-8
PS2O-ex

-10 CPSO

-12

-14

-16
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Generations

Fig. 7. f3, Ackley function.


56 H. Chen, Y. Zhu / Applied Mathematics and Computation 205 (2008) 47–60

2.3

2.2 PS2O-g

2.1 PS2O-l
2 PS2O-ex
1.9 CPSO
Fitness (log)
1.8

1.7

1.6

1.5

1.4

1.3

0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Generations

Fig. 8. f4, Rastrigrin function.

-2

-4

-6
Fitness (log)

-8

-10

-12
PS2O-g
-14
PS2O-l
-16
PS2O-ex
-18
CPSO
-20
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Generations

Fig. 9. f5, Griewank function.

For the unimodal functions, the PS2O-g is able to consistently find the minimum 0 to the Sphere function within 7356
generations. When apply the PS2O-g to Rosenborck function, the average population fitness frequently reach the result of
1.325e-29. Since a result of 40.0 on Rosenbrock is considered good, this is obviously a remarkable performance.
For multimodal functions, the PS2O-l and the PS2O-ex clearly does not suffer from premature convergence. In the 10,000
generations of the test runs, the PS2O-l and the PS2O-ex continues to find better results even after the PS2O-g and the CPSO
seem to have stagnated. It should be mentioned that the PS2O-ex is able to consistently find the minimum 0 of the Griewank
function and the Weierstrass function.

4.4. Experiment 2: simulation of symbiotic coevolution process in PS2O model

In Experiment 2, we present the simulation results of the coevolution process of symbiotic species in our model. To dem-
onstrate the effect of symbiotic coevolution, we run two simulations.
H. Chen, Y. Zhu / Applied Mathematics and Computation 205 (2008) 47–60 57

In the first simulation, we used three benchmark functions (Rosenbrock, Griewank and Weierstrass) as the fitness land-
scape, respectively. Four species (each species possesses 10 members) were randomly initialized to represents the ecosys-
tem. Then the symbiotic coevolution is simulated as these four species adaptively moving over the fitness landscape. In
this ecosystem, the species 1, species 2 and species 3 are symbiotic partners, while species 4 evolved only based on its

-2

-4
PS2O-g
Fitness (log)

-6
PS2O-l
-8 PS2O-ex

-10 CPSO

-12

-14

-16
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Generations

Fig. 10. f6, Weierstrass function.

Table 7
Proportion of experiments reaching criteria/minimum

Proportion of trials reaching criteria Proportion of trials reaching minimum


PS2O-g PS2O-l PS2O-ex CPSO PS2O-g PS2O-l PS2O-ex CPSO
f1 1.00 1.00 1.00 1.00 1.00 0.00 0.00 0.00
f2 1.00 1.00 1.00 1.00 0.00 0.00 0.00 0.00
f3 0.58 1.00 1.00 0.95 0.00 0.00 0.00 0.00
f4 1.00 1.00 1.00 1.00 0.00 0.00 0.00 0.00
f5 1.00 1.00 1.00 0.96 0.40 0.90 1.00 0.30
f6 1.00 1.00 1.00 1.00 0.00 1.00 1.00 0.00

1.6 Species 1
Species 2
Species 3
1.4 Species 4
Fitness (log)

1.2

0.8
Enlarged Graph

0.6
0 200 400 600 800 1000
Generations

Fig. 11. Coevolution on Rosenbrock function.


58 H. Chen, Y. Zhu / Applied Mathematics and Computation 205 (2008) 47–60

Species 1
Species 2
Species 3
0 Species 4

Fitness (log) -5

-10
Enlarged Graph

0 200 400 600 800 1000


Generation

Fig. 12. Coevolution on Griewank function.

2
Species 1
Species 2
1
Species 3
Species 4
0
Fitness (log)

-1

-2

-3 Enlarged Graph

-4
0 200 400 600 800 1000
Generation

Fig. 13. Coevolution on Weierstrass function.

50

40
Extinction (%)

30

20

10

0
0 500 1000 1500 2000
Generation

Fig. 14. Extinction on Ackley function.


H. Chen, Y. Zhu / Applied Mathematics and Computation 205 (2008) 47–60 59

50

40

Extinction (%)
30

20

10

0
0 500 1000 1500 2000
Generation

Fig. 15. Extinction on Rastrigrin function.

own knowledge, i.e., the three symbiotic species evolved according to the rules of the PS2O algorithm (here we employed
PS2O-l), while the species 4 evolved according to the rules of the canonical PSO algorithm. Both algorithms were run
1000 generation on each function.
The simulation results are illustrated in Figs. 11–13. From the results, we can observe that all the symbiotic partners do
not suffer from premature convergence and are able to reach states of higher fitness with greatly faster rate. Just like in nat-
ure, dissimilar species spontaneously establish symbiotic relationships to improve their survivability and adaptability.
In the second simulation, we use two benchmark functions, Ackley and Rastrigrin, as the external stresses imposed on the
ecosystem by the environment. Twenty species (each species possesses five members) were randomly initialized to repre-
sents the ecosystem. We employed the PS2O-ex to evolve on the two fitness landscapes. The simulation was run in 2000
generations. Figs. 14 and 15 show the frequencies of extinctions that occurred during coevolution process on the two bench-
mark functions. From the results, we can observe that there are periods of stasis interspersed with burst of extinctions. This
illustrated mass extinction can be a natural feature of the ecosystem’s dynamics and sometimes is the result of the coevo-
lution of species.

5. Conclusions

In this paper, we present a general model for biological symbiosis in context of coevolving ecosystems. Speciation and
extinction events are also considered in formulating our model. By using this model as a computational metaphor, we have
introduced a novel multi-species optimizer (called PS2O). Each species in PS2O is based on the extension of the control law of
the canonical PSO algorithm. By adding a significant ingredient and extra control to the dynamics of the canonical PSO mod-
el, symbiosis mechanism, species extinction and speciation are simulated in the PS2O model.
A set of six benchmark functions has been used to test three PS2O variants (PS2Os) in comparison with the canonical PSO
algorithm. The simulation results show that, for all the test functions, all PS2Os reach remarkable better fitness values and
have higher success probabilities than the canonical PSO. Since increasing the population size and species number will result
in better and more robust performance of the PS2O algorithms, such an implementation can be used for more difficult mul-
timodal problems.
We also simulated the coevolution process of symbiotic species and the time series of extinction events in the proposed
model. The results capture some important aspects of the dynamics of biological coevolution that some evolutionary biol-
ogists believe takes place in nature.
There are ways to improve our model (e.g., more symbiosis types could be studied, such as commensalisms, and parasit-
ism). Moreover, it remains to be see how practically useful the optimization algorithm are for engineering optimization
problems. These depend on extensive evaluation on many benchmark functions and real-world problems.

Acknowledgements

This work is supported by the National 863 plans projects of China under Grant 2006AA04A124 and the National 863
plans projects of China under Grant 2006AA04A117. The first author would like to thank Prof. John Paddison for reviewing
the manuscript.

References

[1] R.C. Eberchart, J. Kennedy, A new optimizer using particle swarm theory, in: Proceeding of the 6th International Symposium on Micromachine and
Human Science, Nagoya, Japan, 1995, pp. 39–43.
60 H. Chen, Y. Zhu / Applied Mathematics and Computation 205 (2008) 47–60

[2] J. Kennedy, R.C. Eberchart, Particle swarm optimization, in: proceeding of IEEE International Conference on Neural Networks, Piscataway, NJ, 1995, pp.
1942–1948.
[3] M. Dorigo, V. Maniezzo, A. Colorni, Ant system: optimization by a colony of cooperating agents, IEEE Transactions on Systems Man and Cybernetics
(Part B) 26 (1) (1996) 29–41.
[4] M. Dorigo, L. Gambardella, Ant colony system: a cooperative learning approach to the traveling salesman problem, IEEE Transactions on Evolutionary
Computation 1 (1) (1997) 53–66.
[5] R.C. Eberchart, Y. Shi, Particle swarm optimization: developments, applications and resources, in: Proceedings of the IEEE Congress on Evolutionary
Computation, Piscataway, NJ, 2001, pp. 81–86.
[6] J. Kennedy, R.C. Eberchart, Y. Shi, Swarm Intelligence, Morgan Kaufman Publishers, San Francisco, 2001.
[7] S.A. Frank, Models of symbiosis, American Naturalist 150 (1997) 80–99.
[8] M.D. Jason, S.G. Catherine, A.S. Stephen, J.R. Steven, Symbionticism and complex adaptive systems I: implications of having symbiosis occur in nature,
in: Proceedings of the 5th Annual Conference on Evolutionary Programming, Cambridge, 1996, pp. 177–186.
[9] C.R. Marshall, Mass extinction probed, Nature 392 (1998) 17–20.
[10] M. Newman, Simple models of evolution and extinction, IEEE Computing in Science and Engineering (2000) 80–86.
[11] R.A. Watson, J.B. Pollack, How symbiosis can guide evolution, in: Proceedings of the 15th European Conference on Artificial Life, Springer, Dario
Floreano, Jean-Daniel Nicoud, Francesco Mondada, 1999, pp. 29–38.
[12] S. Frank, The origin of synergistic symbiosis, Journal of Theoretical Biology 176 (1995) 403–410.
[13] C.K. Chow, H.T. Tsui, Autonomous agent response learning by a multi-species particle swarm optimization, in: Proceeding of Congress on Evolutionary
Computation, Portland, Oregon, USA, 2004, pp. 778–785.
[14] R. Brits, A.P. Engelbrecht, F. Bergh, A niching particle swarm optimizer, in: Proceedings of the 4th Asia–Pacific Conference on Simulated Evolution and
Learning 2002, pp. 692–696.
[15] K.E. Parsopoulos, M.N. Vrahatis, Modification of the particle swarm optimizer for locating all the global minima, in: Proceedings of the International
Conference on Artificial Neural Networks and Genetic Algorithms (ICANNGA2001), Prague, Czech Republic, 2001, pp. 324–327.
[16] F. Van den Bergh, A.P. Engelbrech, A cooperative approach to particle swarm optimization, IEEE Transactions on Evolutionary Computing 8 (3) (2004)
225–239.
[17] S. Baskar, P.N. Suganthan, A novel concurrent particle swarm optimization, in: Proceedings of the 2004 Congress on Evolutionary Computation, vol. 1,
2004, pp. 792–796.
[18] T. Blackwell, J. Branke, Multi-swarm optimization in dynamic environments, in: Proceedings of Applications of Evolutionary Computing, Lecture Notes
in Computer Science, Springer, vol. 3005, Portugal, 2004, pp. 488–599.
[19] S.A. Kauffman, S. Johnsen, Coevolution to the edge of chaos: coupled fitness landscapes poised states and coevolutionary avalanches, Journal of
Theoretical Biology 149 (1991) 467–505.
[20] M. Clerc, J. Kennedy, The particle swarm: explosion stability and convergence in a multidimensional complex space, IEEE Transactions on Evolutionary
Computation 6 (1) (2002) 58–73.
[21] T. Krink, J.S. Vestertroem, J. Riget, Particle swarm optimization with spatial particle extension, in: Proceedings of the IEEE Congress on Evolutionary
Computation, Honolulu, Hawaii, 2002, pp. 1474–1479.
[22] Y. Shi, R.C. Ebrehart, A modified particle swarm optimizer, in: Proceeding of the 1998 IEEE International Conference on Evolutionary Computation,
Piscataway, NJ, 1998, pp. 69–73.
[23] Y. Shi, R.C. Eberhart, Empirical study of particle swarm optimization, in: Proceedings of the 1999 IEEE Congress on Evolutionary Computation,
Piscataway, NJ, 1999, pp. 1945–1950.
[24] Ben Niu, Yunlong Zhu, XiaoXian He, Xiangping Zeng, Henry Wu, MCPSO: a multi-swarm cooperative particle swarm optimizer, Applied Mathematics
and Computation 185 (2) (2007) 1050–1062.
[25] Jun Zhang, D.S. Huang, Tat-Ming Lok, Michael R. Lyu, A novel adaptive sequential niche technique for multimodal function optimization,
Neurocomputing 69 (16–18) (2006) 2396–2401.
[26] J. Kennedy, Small worlds and mega-minds: effects of neighborhood topology on particle swarm performance, in: Proceedings of the Congress on
Evolutionary Computation, Piscataway, NJ, 1999, pp. 1931–1938.
[27] J. Kennedy, R. Mendes, Population structure and particle swarm performance, in: Proceedings of the 2002 Congress on Evolutionary Computation,
Piscataway, NJ, 2002, pp. 1671–1675.
[28] P.N. Suganthan, Particle swarm optimizer with neighborhood operator, in: Proceedings of the Congress on Evolutionary Computation (CEC 1999),
Piscataway, NJ, 1999, pp. 1958–1962.
[29] X.H. Shi, Y.C. Liang, H.P. Lee, C. Lu, L.M. Wang, An improved GA and a novel PSO-GA-based hybrid algorithm, Information Processing Letters 93 (2005)
255–261.
[30] W.J. Zhang, X.F. Xie, DEPSO: Hybrid particle swarm with differential evolution operator, in: Proceedings of the IEEE International Conference on
Systems, Man and Cybernetics, Washington, DC, USA, 2003, pp. 3816–3821.
[31] B. Ye, C.Z. Zhu, C.X. Guo, Y.J. Cao, Generating extended Fuzzy basis function networks using hybrid algorithm, Lecture Notes in Artificial Intelligence
3613 (2005) 79–88.
[32] S. He, Q.H. Wu, J.Y. Wen, J.R. Saunders, R.C. Paton, A particle swarm optimizer with passive congregation, Biosystems 78 (2004) 35–147.
[33] B. Niu, Y.L. Zhu, X.X. He, Multi-population cooperative particle swarm optimization, in: Proceedings ECAL2005, Lecture Notes in Computer Science No.
3630, 2005, pp. 874–883.

You might also like