P44_2012_IJAIT
-
Upload
irwansyah-ramadhani -
Category
Documents
-
view
212 -
download
0
description
Transcript of P44_2012_IJAIT
![Page 1: P44_2012_IJAIT](https://reader035.fdocuments.in/reader035/viewer/2022070420/5695d0001a28ab9b0290810d/html5/thumbnails/1.jpg)
1st ReadingMay 4, 2012 16:6 WSPC/INSTRUCTION FILE S0218213012400118
International Journal on Artificial Intelligence ToolsVol. 21, No. 3 (2012) 1240011 (20 pages)c© World Scientific Publishing Company
DOI: 10.1142/S0218213012400118
STOCHASTIC STABILITY AND NUMERICAL ANALYSIS
OF TWO NOVEL ALGORITHMS OF THE PSO FAMILY:
PP-GPSO AND RR-GPSO
JUAN LUIS FERNANDEZ-MARTINEZ∗ and ESPERANZA GARCIA-GONZALO†
Mathematic Department, Oviedo University,
Facultad de Ciencias, 33007, Oviedo, Spain∗[email protected]†[email protected]
The PSO algorithm can be physically interpreted as a stochastic damped mass-springsystem: the so-called PSO continuous model. Furthermore, PSO corresponds to a partic-ular discretization of the PSO continuous model. Based on this mechanical analogy wederived in the past a family of PSO-like versions, where the acceleration is discretizedusing a centered scheme and the velocity of the particles can be regressive (GPSO),progressive (CP-GPSO) or centered (CC-GPSO). Although the first and second ordertrajectories of these algorithms are isomorphic, CC-GPSO and CP-GPSO are very dif-ferent from GPSO. In this paper we present two other PSO-like methods: PP-GPSOand RR-GPSO. These algorithms correspond respectively to progressive and regressivediscretizations in acceleration and velocity. PP-PSO has the same velocity update thanGPSO, but the velocities used to update the trajectories are delayed one iteration, thus,PP-PSO acts as a Jacobi system updating positions and velocities at the same time.RR-GPSO is similar to a GPSO with stochastic constriction factor. Both versions havea very different behavior from GPSO and the other family members introduced in thepast: CC-PSO and CP-PSO. RR-PSO seems to have the greatest convergence rate andits good parameter sets can be calculated analytically since they are along a straightline located in the first order stability region. Conversely PP-PSO seems to be a moreexplorative version, although the behavior of these algorithms can be partly problemdependent. Both exhibit a very peculiar behavior, very different from other family mem-bers, and thus they can be called distant PSO relatives. RR-PSO have the greatestconvergence rate of all family members for a wide range of benchmark functions withdifferent numerical complexity in 10, 30 and 50 dimensions. These algorithms have beensuccesfully applied for protein secondary structure prediction and in oil and gas reservoiroptimization.
Keywords: Particle swarm optimization; GPSO; stochastic analysis; convergence.
1. The Generalized PSO (GPSO)
The particle swarm algorithm applied to optimization problems is very simple: in-
dividuals, or particles, are represented by vectors whose length is the number of
degrees of freedom of the optimization problem. To start, a population of parti-
cles is initialized with random positions (x0i ) and velocities (v0
i ). A same objective
1240011-1
![Page 2: P44_2012_IJAIT](https://reader035.fdocuments.in/reader035/viewer/2022070420/5695d0001a28ab9b0290810d/html5/thumbnails/2.jpg)
1st ReadingMay 4, 2012 16:6 WSPC/INSTRUCTION FILE S0218213012400118
J. L. Fernandez-Martınez & E. Garcıa-Gonzalo
function is used to compute the objective value of each particle. As time advances,
the position and velocity of each particle is updated as a function of its objective
function value and of the objective function values of its neighbors. At time-step
k+1, the algorithm updates positions (xk+1i ) and velocities (vk+1
i ) of the individuals
as follows:
vk+1i = ωvk
i + φ1(gk − xk
i ) + φ2(lki − xk
i ) ,
xk+1i = xk
i + vk+1i ,
φ1 = r1ag , φ2 = r2al , r1, r2 ∈ U(0, 1) ω, al, ag ∈ R ,
(1)
where lki is the ith particle’s best position, gk the global best position on the whole
swarm, φ1, φ2 are the random global and local accelerations, and ω is a real constant
called inertia weight. Finally, r1 and r2 are random numbers uniformly distributed
in (0, 1), to weight the global and local acceleration constants, ag and al.
First order stability of this algorithm has been studied by several authors.1,2
Convergence properties of this algorithm and parameter tuning are related to its
exploration capabilities and the stability of second order trajectories.3,4
PSO is the particular case for t = k and ∆t = 1 of the GPSO algorithm:5
v(t+∆t) = (1 − (1− ω)∆t)v(t) + φ1∆t(g(t)− x(t)) + φ2∆t(l(t)− x(t)) ,
x(t+∆t) = x(t) + v(t+∆t)∆t .(2)
This model was derived using a mechanical analogy: a damped mass-spring system
with unit mass, damping factor, 1 − ω, and total stiffness constant, φ = φ1 + φ2,
the so-called PSO continuous model:
x′′(t) + (1− ω)x′(t) + φx(t) = φ1g(t− t0) + φ2l(t− t0) ,
x(0) = x0 ,
x′(0) = v0 .
(3)
In this case x(t) stands for the coordinate trajectory of any particle in the swarm.
In this model particles interact through the local and global attractors, l(t), g(t).
In this model mean particle trajectories oscillate around the point:
o(t) =agg(t− t0) + all(t− t0)
ag + al. (4)
In this model the attractors might be delayed a time t0 with respect to the particle
trajectories.6
S1gpso =
(ω, φ) :∆t− 2
∆t< ω < 1, 0 < φ <
2ω∆t− 2∆t+ 4
∆t2
, (5)
S2gpso =
(ω, φ) : 1− 2
∆t< ω < 1, 0 < φ < φgpso(w,α,∆t)
, (6)
1240011-2
![Page 3: P44_2012_IJAIT](https://reader035.fdocuments.in/reader035/viewer/2022070420/5695d0001a28ab9b0290810d/html5/thumbnails/3.jpg)
1st ReadingMay 4, 2012 16:6 WSPC/INSTRUCTION FILE S0218213012400118
Analysis of Two Novel Algorithms of the PSO Family: PP-GPSO and RR-GPSO
where φ = (ag + al)/2 is the total mean acceleration and φgpso depends on ω,
α, and ∆t and is the analytic expression for the limit hyperbola of second order
stability:
φgpso =12
∆t
(1− ω)(2 + (ω − 1)∆t)
4− 4(ω − 1)∆t+ (α2 − 2α)(2 + (ω − 1)∆t). (7)
α = ag/φ = 2ag/(ag + al) is the ratio between the global acceleration and the total
mean acceleration, and varies in the interval [0, 2]. Low values of α imply for the
same value of φ that the local acceleration is bigger than the global one, and thus,
the algorithm is more explorative. These stability regions do coincide for ∆t = 1
with those shown in previous analyses for the PSO case.1,5,7
Figure 1 shows for the PSO case the first and second order stability regions and
their corresponding spectral radii. The spectral radii are related to the attenuation
of the first and second order trajectories. In the PSO case, the first order spectral
radius is zero in (ω, φ) = (0, 1). The first order stability zone (S1gpso) only depends
on (ω, φ), while the second order stability region (S2gpso) depends on (ω, ag, al). Also,
the second order stability region is embedded in the first order stability region, and
depends symmetrically on α, reaching its maximum size when α = 1 (al = ag).
Good parameter sets are close to the upper limit of second order stability5 as we
can observed in figure 2 that shows for the Griewank, Rosenbrock, Rastrigin and
De Jong-f4 functions the median logarithmic error after 50 simulations for a lattice
of (ω, φ) points located on the GPSO first stability region. Fernandez Martınez
and Garcıa Gonzalo6 also derived CP-GPSO and CC-GPSO which correspond to
two different discretizations of the PSO continuous model when the acceleration is
centered and the velocity is either progressive or centered. These two versions are
very different from PSO: CP-GPSO seems to have a more explorative character,
while CC-GPSO seems to locate faster the global minimum using two consecutive
centers of attraction.
ω
φ
(a) GPSO First Order Spectral Radius
−1 −0.5 0 0.5 10
0.5
1
1.5
2
2.5
3
3.5
4
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
ω
φ
(b) GPSO Second Order Spectral Radius
−1 −0.5 0 0.5 10
0.5
1
1.5
2
2.5
3
3.5
4
0.4
0.5
0.6
0.7
0.8
0.9
1
Fig. 1. PSO: First and second order stability region and corresponding spectral radii.
1240011-3
![Page 4: P44_2012_IJAIT](https://reader035.fdocuments.in/reader035/viewer/2022070420/5695d0001a28ab9b0290810d/html5/thumbnails/4.jpg)
1st ReadingMay 4, 2012 16:6 WSPC/INSTRUCTION FILE S0218213012400118
J. L. Fernandez-Martınez & E. Garcıa-Gonzalo
ω
φ
(a) Griewank
−1 −0.5 0 0.5 10
0.5
1
1.5
2
2.5
3
3.5
4
−0.5
0
0.5
1
1.5
2
2.5
3
ω
φ
(b) Rosenbrock
−1 −0.5 0 0.5 10
0.5
1
1.5
2
2.5
3
3.5
4
3
4
5
6
7
8
ω
φ
(c) Rastrigin
−1 −0.5 0 0.5 10
0.5
1
1.5
2
2.5
3
3.5
4
2
2.1
2.2
2.3
2.4
2.5
2.6
2.7
2.8
ω
φ
(d) De Jong f4
−1 −0.5 0 0.5 10
0.5
1
1.5
2
2.5
3
3.5
4
−2
0
2
4
6
Fig. 2. PSO: Mean error contourplot (in log10 scale) for the Griewank, Rosenbrock, Rastriginand De Jong f4 functions in 50 dimensions.
Full stochastic analysis of the PSO continuous and discrete models (GPSO) has
been performed by Fernandez Martınez and Garcıa Gonzalo.3 This analysis served
to analyze the GPSO second order trajectories, to show the convergence of GPSO
to the continuous PSO model as the discretization time step goes to zero and to
analyze the role of the oscillation center on the first and second order continuous and
discrete dynamical systems. This analysis also shed light about PSO convergence
for a wide class of benchmark functions when the PSO parameters are selected close
to the upper border of the second order stability region.
In this paper we present two other algorithms belonging to PSO extended family:
PP-GPSO and RR-GPSO. These algorithms correspond respectively to progressive
and regressive discretizations in acceleration and velocity. PP-GPSO has the same
velocity update than GPSO, but the velocities used to update the trajectories are
delayed one iteration. RR-GPSO can be interpreted as a GPSO with a stochastic
constriction factor that depends on the iterations. In both cases we derive their first
and second order stability regions.
1240011-4
![Page 5: P44_2012_IJAIT](https://reader035.fdocuments.in/reader035/viewer/2022070420/5695d0001a28ab9b0290810d/html5/thumbnails/5.jpg)
1st ReadingMay 4, 2012 16:6 WSPC/INSTRUCTION FILE S0218213012400118
Analysis of Two Novel Algorithms of the PSO Family: PP-GPSO and RR-GPSO
Finally we perform a numerical comparison between the different members of
the family for a set of benchmark functions in 50 dimensions. RR-PSO shows (at
least in these cases) convergence rates similar or even greater than PSO, while PP-
PSO is a more explorative version. This is coherent with the fact that PP-PSO acts
as a Jacobi system updating positions and velocities at the same time. These results
are very promising, although they might be different for other kind of benchmark
functions exhibiting other numerical difficulties. Nevertheless these results are very
important in the application of PSO to inverse problems, because due to the non-
convex character of these kind of problems, it is not only important to achieve very
low misfits but also to explore the space of possible solutions that are compatible
with the observed data and the prior information that is at disposal.8,9 RR-PSO
has been applied to the reservoir history matching inverse problems in oil and gas
and has given very impressive results.10 These algorithms have been also succesfully
applied for protein secondary structure prediction.11,12
2. The PP-GPSO Algorithm
Let us use progressive discretizations in acceleration and in velocity to approximate
the PSO continuous model (3):
x′(t) ≃ x(t+∆t)− x(t)
∆t, x′′(t) ≃ x(t+ 2∆t)− 2x(t+∆t) + x(t)
∆t2, (8)
The acceleration in this case corresponds to a progressive discretization in velocity
x′′(t) ≃ x′(t+∆t)− x′(t)
∆t. (9)
The following relationships apply:
x(t+∆t) = x(t) + v(t)∆t ,
v(t+∆t)− v(t)
∆t+ (1− ω)v(t) = φ1(g(t− t0)− x(t)) + φ2(l(t− t0)− x(t)) .
(10)
Adopting t0 = 0 we arrive at:
v(t+∆t) = (1 − (1− ω)∆t)v(t) + φ1∆t(g(t)− x(t)) + φ2∆t(l(t)− x(t)) ,
x(t+∆t) = x(t) + v(t)∆t ,(11)
which has the same expression for the velocity that the GPSO. The main difference
is that the velocity used to update the trajectory is v(t) instead of v(t +∆t) that
is used in the GPSO. This fact will provoke PP-GPSO to be more explorative than
GPSO and having a lower convergence rate. This fact has already been pointed in
the CP-PSO case.6 In the next section we analyze which are the similarities and
differences between both algorithms regarding their first and second order stability
regions.
1240011-5
![Page 6: P44_2012_IJAIT](https://reader035.fdocuments.in/reader035/viewer/2022070420/5695d0001a28ab9b0290810d/html5/thumbnails/6.jpg)
1st ReadingMay 4, 2012 16:6 WSPC/INSTRUCTION FILE S0218213012400118
J. L. Fernandez-Martınez & E. Garcıa-Gonzalo
2.1. The first and second order stability regions
The following stochastic second order difference equation is obtained for PP-GPSO
algorithm:
x(t +∆t)−Appx(t) −Bppx(t −∆t) = Cpp(t−∆t) ,
App = 2− (1 − ω)∆t ,
Bpp = (1 − ω)∆t− 1− φ∆t2 ,
Cpp(t−∆t) = (φ1g(t−∆t) + φ2l(t−∆t))∆t2 .
(12)
The first order moments satisfy the affine dynamical system:
µ(t+∆t) = Appµ µ(t) + bpp
µ (t) , (13)
where
Appµ =
(
A1 E(B1)
1 0
)
, bppµ (t) =
(
φ∆t2E(o(t −∆t))
0
)
, (14)
and
E(o(t)) =agE(g(t)) + alE(l(t))
ag + al. (15)
The first order stability region of the PP-GPSO is:
S1pp =
(ω, φ) :∆t− 4
∆t< ω < 1, 0 < 2
(1− ω)∆t− 2
∆t2< φ <
(1 − ω)
∆t
. (16)
The curve separating real and complex eigenvalues ofAppµ in the first stability region
is independent of ∆t:
φ =1
4(1− 2ω + ω2) . (17)
The spectral radius is zero on the point:
(ω, φ) =
(
1− 2
∆t,
1
∆t2
)
. (18)
The non-centered second order moments satisfy the following second order affine
dynamical system:
r2(t+∆t) = Appσ r2(t) + bpp
r(t) , (19)
where
Appσ =
A2pp 2AppE(Bpp) E(B2
pp)
App E(Bpp) 0
1 0 0
, (20)
1240011-6
![Page 7: P44_2012_IJAIT](https://reader035.fdocuments.in/reader035/viewer/2022070420/5695d0001a28ab9b0290810d/html5/thumbnails/7.jpg)
1st ReadingMay 4, 2012 16:6 WSPC/INSTRUCTION FILE S0218213012400118
Analysis of Two Novel Algorithms of the PSO Family: PP-GPSO and RR-GPSO
and
bppr(t)1 = E(C2
1 (t−∆t)) + 2A1E(C1(t−∆t)x(t))
+ 2E(B1)E(C1(t−∆t)x(t −∆t)) ,
bppr(t)2 = E(C1(t−∆t)x(t)) ,
bppr(t)3 = 0 .
(21)
The second order stability region S2pp is made up by the pairs (ω, φ):
[
ωv < ω < 1− 2
∆t, φ−
pp(ω, α,∆t) < φ < φ+pp(ω, α,∆t)
]
∪[
1− 2
∆t< ω < 1, 0 < φ < φ+
pp(ω, α,∆t)
]
(22)
where φ = φ−
pp(ω, α,∆t), φ = φ+pp(ω, α,∆t) are the upper limits hyperbolas of
second order stability and ωv > 1− (4/∆t). This curve has a complicate analytical
equation. For example for α = 1 and ∆t = 1 its equation is given by:
ωv = −1.73 ,
φ+pp(ω, 1, 1) =
1
14
(
− 7− 19ω +√385 + 266ω + 25ω3
)
,
φ−
pp(ω, 1, 1) =1
14
(
− 7− 19ω −√385 + 266ω + 25ω3
)
.
(23)
As for the GPSO case, the region S2pp is embedded in S1
pp.
PP-PSO is the particular case where the time step is ∆t = 1. Figure 3 shows
the first and second order stability regions of the PP-PSO case (∆t = 1) with
the associated spectral radii. For the case of second order region the parameter α
has been set to 1 in this case. As it can be observed both regions of stability are
bounded. Generally speaking both stability regions are a tilted versions of the PSO
regions.
Figure 4 shows the correspondence between the homogeneous first order trajec-
tories as defined in the paper.2 The fact that these regions are linearly isomorphic
does not mean that both algorithms are the same, because of their stochastic char-
acter and how they update the force term (the oscillation center o(t) in PP-PSO is
delayed as shown in the Cpp(t−∆t) constant). Finally figure 5 shows for PP-PSO
the logarithmic error for the Griewank, Rosenbrock, Rastrigin and De Jong-f4 case
in 50 dimensions. Compared to figure 2 it can be observed that PP-PSO provides
greater misfits than the PSO for all the functions. This has to be with the fact
that PP-PSO updates at the same time the velocities and positions of the particles.
1240011-7
![Page 8: P44_2012_IJAIT](https://reader035.fdocuments.in/reader035/viewer/2022070420/5695d0001a28ab9b0290810d/html5/thumbnails/8.jpg)
1st ReadingMay 4, 2012 16:6 WSPC/INSTRUCTION FILE S0218213012400118
J. L. Fernandez-Martınez & E. Garcıa-Gonzalo
ω
φ
(a) PP−PSO First order spectral radius
−3 −2 −1 0 10
0.5
1
1.5
2
2.5
3
3.5
4
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
ω
φ
(b) PP−PSO Second order spectral radius
−3 −2 −1 00
0.5
1
1.5
2
2.5
3
3.5
0.4
0.5
0.6
0.7
0.8
0.9
1
Fig. 3. PP-PSO: First and second order stability regions and corresponding spectral radii.
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
v v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
+
+
+
+
+
+
+
+
+
+
+
+
+
+ + + + + + +
x
x
x
x
x
x
x
x
x x
xx
xx
x
x
x
x
x
x
-1.0 -0.5 0.0 0.5 1.00
1
2
3
4
Ω
Φ-
PSO
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
ov v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v+
+
+
+
+
+
+
+
+
+
+
+
+
+ + + + + ++
x
x
x
x
x
x
x
x
x x
xx
xx
x
x
x
x
x
x
-3 -2 -1 0 10
1
2
3
4
Ω
Φ-
PP-PSO
Fig. 4. Correspondence between the homogeneous first order trajectories of PSO and PP-PSO.
Also it can be observed that the algorithm does not converge for ω < 0, and the
good parameter sets are in the complex region, that can be seen in figure 4, close to
the limit of second order stability and close to φ = 0. These results can be partially
altered by clamping the velocities or by varying the time step, ∆t. In conclusion it is
expected for the PP-PSO a more explorative behavior (and thus lower convergence
rates) than for the PSO case. This analysis also shows that how the PSO algorithm
was proposed is a kind of coincidence because if in the velocity update v(t) had been
used instead of v(t+1) the results of this algorithm would not be so impressive as in
the PSO case and maybe today we would not be writing this paper. This illustrates
the importance of the velocity update in the PSO convergence.
1240011-8
![Page 9: P44_2012_IJAIT](https://reader035.fdocuments.in/reader035/viewer/2022070420/5695d0001a28ab9b0290810d/html5/thumbnails/9.jpg)
1st ReadingMay 4, 2012 16:6 WSPC/INSTRUCTION FILE S0218213012400118
Analysis of Two Novel Algorithms of the PSO Family: PP-GPSO and RR-GPSO
ω
φ
(a) Griewank
−3 −2 −1 00
1
2
3
4
0.5
1
1.5
2
2.5
3
ω
φ
(b) Rosenbrock
−3 −2 −1 00
1
2
3
4
4.5
5
5.5
6
6.5
7
7.5
8
8.5
ω
φ
(c) Rastrigin
−3 −2 −1 00
1
2
3
4
2.1
2.2
2.3
2.4
2.5
2.6
2.7
2.8
ω
φ
(d) De Jong f4
−3 −2 −1 00
1
2
3
4
3
4
5
6
7
Fig. 5. PP-PSO: Mean error contourplot (in log10 scale) for the Griewank, Rosenbrock, Rastriginand De Jong f4 functions in 50 dimensions.
3. The RR-GPSO Algorithm
Let us adopt a regressive discretization in acceleration and in velocity in time t ∈ R,
in order to discretize model (3) :
x′(t) ≃ x(t)− x(t−∆t)
∆t,
x′′(t) ≃ x(t)− 2x(t−∆t) + x(t− 2∆t)
∆t2=
x′(t)− x′(t−∆t)
∆t.
(24)
The following relationships apply:
x(t) = x(t−∆t) + v(t)∆t ,
v(t) − v(t−∆t)
∆t+ (1 − ω)v(t) + φ(x(t −∆t) + v(t)∆t)
= φ1g(t− t0) + φ2l(t− t0) ,
(25)
1240011-9
![Page 10: P44_2012_IJAIT](https://reader035.fdocuments.in/reader035/viewer/2022070420/5695d0001a28ab9b0290810d/html5/thumbnails/10.jpg)
1st ReadingMay 4, 2012 16:6 WSPC/INSTRUCTION FILE S0218213012400118
J. L. Fernandez-Martınez & E. Garcıa-Gonzalo
v(t) =v(t−∆t) + φ1∆t(g(t− t0)− x(t−∆t))
1 + (1− ω)∆t+ φ∆t2
+φ2∆t(l(t− t0)− x(t−∆t))
1 + (1− ω)∆t+ φ∆t2. (26)
RR-GPSO algorithm can be written as:
x(t +∆t) = x(t) + v(t+∆t)∆t ,
v(t+∆t) =v(t) + φ1∆t(g(t+∆t− t0)− x(t))
1 + (1− ω)∆t+ φ∆t2
+φ2∆t(l(t+∆t− t0)− x(t))
1 + (1 − ω)∆t+ φ∆t2.
(27)
The natural choice for t0 is ∆t. Thus, the RR-GPSO algorithm with delay one
becomes:
v(t+∆t) =v(t) + φ1∆t(g(t)− x(t)) + φ2∆t(l(t)− x(t))
1 + (1− ω)∆t+ φ∆t2,
x(t +∆t) = x(t) + v(t+∆t)∆t , t,∆t ∈ R ,
x(0) = x0 , v(0) = v0 .
(28)
RR-PSO with delay one is a particular case of (28) for a unit time step, ∆t = 1.
RR-PSO is a PSO-like algorithm where the parameter
A(ω, φ,∆t) =1
1 + (1− ω)∆t+φ∆t2(29)
could be interpreted as a constriction factor similar to the introduced by Clerc and
Kennedy (2002). The fact that there is a delay one on the parameter t0 causes that
there will be only a correspondence between the homogeneous trajectories (without
taking into account the force term) of RR-PSO and PSO. This fact has been also
outlined for other PSO family members.6
3.1. The first and second order stability regions
The following stochastic second order difference equation is obtained for the RR-
GPSO algorithm:
x(t+∆t)−Arrx(t)−Brrx(t−∆t) = Crr(t+∆t) , (30)
where
Arr =2 + (1− ω)∆t
1 + (1− ω)∆t+ φ∆t2, Brr =
−1
1 + (1− ω)∆t+ φ∆t2∈ R , (31)
1240011-10
![Page 11: P44_2012_IJAIT](https://reader035.fdocuments.in/reader035/viewer/2022070420/5695d0001a28ab9b0290810d/html5/thumbnails/11.jpg)
1st ReadingMay 4, 2012 16:6 WSPC/INSTRUCTION FILE S0218213012400118
Analysis of Two Novel Algorithms of the PSO Family: PP-GPSO and RR-GPSO
and
Crr(t+∆t) =φ1g(t+∆t− t0) + φ2l(t+∆t− t0)
1 + (1− ω)∆t+ φ∆t2∆t2 . (32)
The first order affine system describing the RR-GPSO mean trajectories is:
µ(t+∆t) = ARRµ µ(t) + bRR
µ (t) , (33)
where
ARRµ =
(
E(Arr) E(Brr)
1 0
)
, bRRµ (t) =
φE(o(t+∆t− t0))
1 + (1− ω)∆t+ φ∆t2∆t2
0
. (34)
The first order stability region of RR-GPSO is composed of two different disjoint
zones D1 and D2 described by S1RR−gpso = (ω, φ) : D1 ∪D2, where:
D1 =
[ω < 1, φ > 0] ∪[
1 < ω < 1 +4
∆t, φ >
ω − 1
∆t
]
∪[
ω > 1 +4
∆t, φ >
2
∆t2((ω − 1)∆t− 2)
]
, (35)
and
D2 =
[
ω < 1 +2
∆t, φ <
2
∆t2((ω − 1)∆t− 2)
]
∪[
ω > 1 +2
∆t, φ < 0
]
. (36)
The parabola separating the real and complex eigenvalues of ARRµ is the same as in
PP-PSO case and does not depend on ∆t:
φ =1
4(1− 2ω + ω2) . (37)
For ∆t = 1, this first order region of stability becomes:
S1RR = (ω, φ) : D1 ∪D2 ,
D1 =
[ω < 1, φ > 0] ∪ [1 < ω < 5, φ > ω − 1]
∪[ω > 5, φ > 2(ω − 3)]
,
D2 = [ω < 3, φ < 2(ω − 3)] ∪ [ω > 3, φ < 0].
(38)
The non-centered second order moments fulfill the following second order affine
dynamical system:
r2(t+∆t) = ARRσ r2(t) + bRR
r(t) , (39)
1240011-11
![Page 12: P44_2012_IJAIT](https://reader035.fdocuments.in/reader035/viewer/2022070420/5695d0001a28ab9b0290810d/html5/thumbnails/12.jpg)
1st ReadingMay 4, 2012 16:6 WSPC/INSTRUCTION FILE S0218213012400118
J. L. Fernandez-Martınez & E. Garcıa-Gonzalo
ω
φ
(a) RR−PSO First order spectral radius
−10 −5 0 5−10
−8
−6
−4
−2
0
2
4
6
8
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
ω
φ
(b) RR−PSO Second order spectral radius
−10 −5 0 5−10
−8
−6
−4
−2
0
2
4
6
8
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Fig. 6. RR-PSO: First and second order stability regions and corresponding spectral radii.
where
ARRσ =
E(A2rr) 2E(ArrBrr) E(B2
rr)
(Arr) E(Brr) 0
1 0 0
, (40)
and
brrr(t)1 = E(C2
1 (t+∆t)) + 2E(AC1(t+∆t)x(t))
+ 2E(B1C1(t+∆t)x(t −∆t)) ,
brrr(t)2 = E(C1(t+∆t)x(t)) ,
brrr(t)3 = 0 .
(41)
The eigenvalues analysis of second order iteration matrix ARRσ allows to determine
the RR-PSO second order stochastic stability region.
Figure 6 shows for ∆t = 1 (RR-PSO case) and α = 1 (ag = al), the first
and second order stability regions with the corresponding first and second order
spectral radii. Both regions of stability are unbounded. Also in both cases the first
and second order spectral radii are zero at the infinity: (ω, φ) = (−∞,+∞) and
(ω, φ) = (+∞,−∞). Figure 7 shows the correspondence between the homogeneous
trajectories of RR-PSO and PSO. Figure 8 shows for RR-PSO the logarithmic error
for the Griewank, Rosenbrock, Rastrigin and de Jong-f4 case in 50 dimensions.
Compared to figure 2 it can be observed that RR-PSO provides lower misfits than
the PSO for all the benchmark functions. The difference in some cases is very
significative (Rastrigin and de Jong-f4). These results have been also found for the
same functions in 10 and 30 dimensions.
The good parameters sets for RR-PSO seem to be concentrated around the line
of equation φ = 3(ω−(3/2)),mainly for inertia values greater than 2. This line is the
same for both functions and seems to be invariant when the number of parameters
1240011-12
![Page 13: P44_2012_IJAIT](https://reader035.fdocuments.in/reader035/viewer/2022070420/5695d0001a28ab9b0290810d/html5/thumbnails/13.jpg)
1st ReadingMay 4, 2012 16:6 WSPC/INSTRUCTION FILE S0218213012400118
Analysis of Two Novel Algorithms of the PSO Family: PP-GPSO and RR-GPSO
ooooo
oooooooooo
oooooooooooooo
ooooooooooooooooo
ooooooooooooooooooo
ooooooooooooooooooooo
ooooooooooooooooooooooo
ooooooooooooooooooooooooo
oooooooooooooooooooooooooo
oooooooooooooooooooooooooooo
ooooooooooooooooooooooooooooo
oooooooooooooooooooooooooooooo
oooooooooooooooooooooooooooooooo
ooooooooooooooooooooooooooooooooo
oooooooooooooooooooooooooooooooooo
ooooooooooooooooooooooooooooooooooo
ooooooooooooooooooooooooooooooooooooo
oooooooooooooooooooooooooooooooooooooo
ooooooooooooooooooooooooooooooooooooooo
oooooooooooooooooooooooooooooooooooooooo
v vv
vvv
vvvv
vvvvv
vvvvvv
vvvvvvv
vvvvvvvv
vvvvvvvvv
vvvvvvvvvv
vvvvvvvvvvv
vvvvvvvvvvvv
vvvvvvvvvvvvv
vvvvvvvvvvvvvv
vvvvvvvvvvvvvvv
vvvvvvvvvvvvvvvv
vvvvvvvvvvvvvvvvv
vvvvvvvvvvvvvvvvvv
vvvvvvvvvvvvvvvvvvv
vvvvvvvvvvvvvvvvvvvv
vvvvvvvvvvvvvvvvvvvvv
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+ + + + + + + + + + +
xxxxxxxxx
xxxxxx
xxxxx
xxxx
xxx x
xx
xx x
x xx x
x xx
xx
xx
xx
xx
-1.0 -0.5 0.0 0.5 1.00
1
2
3
4
Ω
Φ-
PSO
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
oo
oo
oo
oo
oo
oo
oo
oo
oo
o
oo
oo
oo
oo
oo
oo
oo
oo
oo
oo
o
oo
oo
oo
oo
oo
oo
oo
oo
oo
oo
oo
o
oo
oo
oo
oo
oo
oo
oo
oo
oo
oo
oo
oo
o
oooooooooooooooooooooooooo
oooooooooooooooooooooooooooo
ooooooooooooooooooooooooooooo
oooooooooooooooooooooooooooooo
oooo
oooo
oooo
oooo
oooo
oooo
oooo
oooo
oooo
oooo
oooo
oooo
oooo
oooo
oooo
oooo
o
oooo
oooo
oooo
oooo
oooo
oooo
oooo
oooo
oo
oooo
oooo
oooo
oooo
oooo
oooo
oooo
oooo
ooo
oooooo
oooooo
oooooo
oooooo
oooooo
oooooo
o
oooooo
oooooo
oooooo
oooooo
oooooo
oooooo
oo
oooooo
oooooo
oooooo
oooooo
oooooo
oooooo
ooo
oooooo
oooooo
oooooo
oooooo
oooooo
oooooo
ooo
o vvvvvv vvvvvv vvvvvvvvvv
vvvvvvv
vvvvvvvv
vvvvvvvvv
vvvvvvvvvv
vvvvvvvvvvv
vvvvvvvvvvvv
vv
vv
vv
vv
vv
vv
v
vv
vv
vv
vv
vv
vv
vv
vv
vv
vv
vv
vv
vv
vv
v
vv
vv
vv
vv
vv
vv
vv
vv
vv
vv
vv
vv
vv
vv
vv
vv
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v v
v
+
+
+
+
+
+
+
+
+
+
+
+
+++
+++
+++++++++++++++++
xx
xx
x
xx
xx
xx
xxx
xxxxxxxx
x
-5 0 5 10-10
-5
0
5
10
Ω
Φ-
RR-PSO
Fig. 7. Correspondence between the homogeneous first order trajectories of PSO and RR-PSO.
Griewank
ω
φ
0 2 4 6 80
2
4
6
8
10
12
−2
−1
0
1
2
3Rosenbrock
ω
φ
0 2 4 6 80
2
4
6
8
10
12
3
4
5
6
7
8
Rastrigin
ω
φ
0 2 4 6 80
2
4
6
8
10
12
0.5
1
1.5
2
2.5de Jong−f4
ω
φ
0 2 4 6 80
2
4
6
8
10
12
−6
−4
−2
0
2
4
6
Fig. 8. RR-PSO: Mean error contourplot (in log10 scale) for the Griewank, Rosenbrock, Rastriginand De Jong f4 functions in 50 dimensions.
1240011-13
![Page 14: P44_2012_IJAIT](https://reader035.fdocuments.in/reader035/viewer/2022070420/5695d0001a28ab9b0290810d/html5/thumbnails/14.jpg)
1st ReadingMay 4, 2012 16:6 WSPC/INSTRUCTION FILE S0218213012400118
J. L. Fernandez-Martınez & E. Garcıa-Gonzalo
ω
φ
(a) RR−PSO second order spectral radius
−10 −5 0 5−10
−5
0
5
10
15
20
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
ω
φ
(b) RR−PSO second order trajectory frequency
−10 −5 0 5 10−10
−5
0
5
10
15
20
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
Fig. 9. RR-PSO: second order spectral radius and frequency of the second order moments.
of the optimization function increases. This result is very different from the ones
shown for the other version, since the good parameters are not in relation with the
second order stability upper border. Figure 9 shows the position of this line with
respect to the RR-PSO second order spectral radius and the frequency of the second
order moments trajectories as shown in the paper6 for PSO. This line is located in
a zone of medium attenuation and very high frequency of trajectories. This last
property might be the cause of its good properties, since this feature allows for a
very efficient and explorative search around the oscillation center of each particle
in the swarm.
4. Correspondence Rules
Since all the PSO algorithms presented in this paper come from different finite dif-
ference schemes of the same continuous model, there should be a correspondence
between discrete trajectories for the different PSO versions, that are deduced iden-
tifying the coefficients in their corresponding difference equations, as we have shown
in the paper.6 The corresponding rules to make these algorithms corresponding in
their homogeneous parts, as wPSO = wPP +∆tφPP , are:
wPSO =wRR −∆tφRR + (1− wRR)∆t+∆t2φRR
1 + (1− wRR)∆t+∆t2φRR
,
φPSO = φPP =φRR
1 + (1− wRR)∆t+∆t2φRR
.
(42)
These relationships have been used in figures 4 and 7 to establish the correspondence
of the first trajectories between PP-GPSO, RR-GPSO and GPSO.
1240011-14
![Page 15: P44_2012_IJAIT](https://reader035.fdocuments.in/reader035/viewer/2022070420/5695d0001a28ab9b0290810d/html5/thumbnails/15.jpg)
1st ReadingMay 4, 2012 16:6 WSPC/INSTRUCTION FILE S0218213012400118
Analysis of Two Novel Algorithms of the PSO Family: PP-GPSO and RR-GPSO
It has been shown in the paper6 that these correspondences make the algorithms
linearly isomorphic, but when applied to real optimization problems they show
different performances or two main reasons:
(1) The introduction of a delay parameter increases the exploration and makes the
algorithms different in the way they update the force terms. This is the case for
RR-GPSO.
(2) The different way these algorithms update the positions and velocities of the
particles. This is the case for PP-GPSO.
To show that these algorithms are not equivalent, we perform the following
numerical experiment. We considered the Clerc and Kennedy’s point1 for PSO and
we calculated for the other members of the family the corresponding points using
the formulae (42) and the corresponding rules stated in the paper.6 We compute the
median misfit of 100 different simulations after 200 iterations. For these simulations
we have used a swarm of 20 particles for 10 dimensions, 40 particles for 30 dimensions
and 100 particles for 50 dimensions, as it is recommended on the benchmark tests.13
The search spaces for these functions are:
Ackley [−32, 32]D
Alpine [−32, 32]D
deJong-f4 [−32, 32]D
Griewank [−600, 600]D
Sphere [−100, 100]D
Rastrigin [−5.12, 5.12]D
Rosenbrock [−30, 30]D
Table 1 shows the results of these simulations. As expected convergence rate is
very different for all the family members. We have already shown this point in the
paper6 for GPSO, CC-GPSO and CP-GPSO.
5. Numerical Experiments
Although we have shown some numerical results in figures 2, 5 and 8, in the next
section we show some additional comparisons between all the members of the family.
Table 2 shows the median convergence curves after 100 different simulations and 200
iterations, for different benchmark functions. For PSO we have considered the Clerc
and Kennedy’s point (ω, φ) = (0.729, 1.494) that has a very good performance. For
CC-PSO, CP-PSO and PP-PSO we have considered the following points selected
from the points that are located on low misfit region of the corresponding algorithms:
• CC: (ω, φ) = (0.74, 1.95),
• PP: (ω, φ) = (0.88, 0.1),
• CP: (ω, φ) = (0.96, 0.4).
1240011-15
![Page 16: P44_2012_IJAIT](https://reader035.fdocuments.in/reader035/viewer/2022070420/5695d0001a28ab9b0290810d/html5/thumbnails/16.jpg)
1st ReadingMay 4, 2012 16:6 WSPC/INSTRUCTION FILE S0218213012400118
J. L. Fernandez-Martınez & E. Garcıa-Gonzalo
Table 1. Median errors for different benchmark functions and algorithms using some Clerc
and Kennedy’s point and the corresponding (ω, φ) points.
Function dim PSO RR PP CC CP
10 1.1e-03 6.9e+00 1.9e+01 1.6e+00 1.3e+00
Ackley 30 2.0e+00 1.4e+01 1.9e+01 8.6e+00 2.0e+01
50 3.3e+00 1.6e+01 1.9e+01 1.1e+01 2.0e+01
10 5.3e-03 6.4e+00 4.7e+01 9.6e-02 3.4e+00
Alpine 30 4.9e+00 6.4e+01 2.1e+02 1.2e+01 1.8e+02
50 1.8e+01 1.2e+02 3.6e+02 3.2e+01 3.4e+02
10 6.5e-12 8.3e+02 2.9e+06 3.1e-28 1.4e+01
deJong-f4 30 1.6e+00 9.7e+05 5.4e+07 1.7e+02 5.1e+07
50 3.7e+02 6.2e+06 1.4e+08 9.4e+03 1.4e+08
10 1.0e-01 5.2e+00 1.3e+02 1.2e-01 8.1e-01
Griewank 30 9.3e-01 9.2e+01 6.5e+02 1.8e+00 6.4e+02
50 1.2e+00 2.2e+02 1.1e+03 7.5e+00 1.1e+03
10 3.5e-06 2.3e+02 1.5e+04 7.6e-14 1.1e+00
Sphere 30 1.6e+00 7.5e+03 7.1e+04 7.7e+01 6.9e+04
50 2.9e+01 1.9e+04 1.2e+05 6.9e+02 1.1e+05
10 7.3e+00 2.9e+01 1.0e+02 1.9e+01 4.7e+01
Rastrigin 30 5.1e+01 2.1e+02 4.3e+02 1.0e+02 4.2e+02
50 1.0e+02 4.0e+02 7.4e+02 1.9e+02 7.4e+02
10 6.8e+00 9.2e+03 3.8e+07 8.9e+00 5.3e+02
Rosenbrock 30 2.9e+02 5.2e+06 2.9e+08 2.8e+03 2.7e+08
50 3.2e+03 2.2e+07 4.8e+08 3.4e+04 4.6e+0
For RR-PSO we have used two points named P1 : (ω, φ) = (3.9, 6.97) and P2 :
(ω, φ) = (1.8, 0.34). Point P2 seems not to outperform (see figure 8) for functions of
higher numerical complexity such as Rosenbrock and Griewank, while point P1seems
to work better for the test of benchmark functions with lower numerical complexity.
It is possible to observe that RR-PSO is the algorithm that performs the best for
all the benchmark functions. Point P1 works better than P2 for Griewank and
Rosenbrock. For the other functions with lower complexity the results obtained
by RR-PSO using P2 are very impressive. The RR-PSO algorithm is able to find
the global minimum within 50 iterations. Figure 10 shows the median convergence
curves for 1000 runs for different benchmark functions in 50 dimensions. In this
case to run RR-PSO we have adopted the point P2. It is possible to observe that
for most of the functions the RR-PSO curve stops suddenly, meaning that the
algorithm has found the global minimum. Table 2 (first column) shows the results
that we have obtained using the algorithm Standard-PSO-200714 and the program
developed by Birge.15 The Standard-PSO-2007 algorithm provides very bad results
compared to the rest of the algorithms (Table 2) when only 200 iterations are used,
1240011-16
![Page 17: P44_2012_IJAIT](https://reader035.fdocuments.in/reader035/viewer/2022070420/5695d0001a28ab9b0290810d/html5/thumbnails/17.jpg)
1st ReadingMay 4, 2012 16:6 WSPC/INSTRUCTION FILE S0218213012400118
Analysis of Two Novel Algorithms of the PSO Family: PP-GPSO and RR-GPSO
Table
2.
Med
ianerrors
fordifferen
tben
chmark
functionsandalgorithmsusingsomeperform
ing(ω
,φ)points.
Function
dim
PSO-S
tdPSO-B
irge
RR(P
1)
RR(P
2)
PP
CC
CP
10
1.2e+
01
1.3e-03
2.3e-07
-8.9e-16
5.3e+
00
1.6e-05
2.8e+
00
Ackley
30
1.8e+
01
2.1e+
00
7.7e-02
-8.9e-16
7.6e+
00
6.1e+
00
4.9e+
00
50
1.8e+
01
3.2e+
00
5.6e-01
-8.9e-16
7.5e+
00
8.1e+
00
5.1e+
00
10
2.0e+
01
4.5e-03
3.8e-07
0.0e+
00
7.4e+
00
6.2e-03
6.3e+
00
Alpine
30
1.1e+
02
5.0e+
00
3.4e-02
0.0e+
00
5.5e+
01
5.7e+
00
7.4e+
01
50
2.1e+
02
1.7e+
01
3.9e-01
0.0e+
00
9.8e+
01
1.6e+
01
1.3e+
02
10
2.7e+
04
6.2e-12
1.6e-23
0.0e+
00
1.9e+
02
5.4e-21
2.9e+
00
deJong-f4
30
4.1e+
06
1.9e+
00
4.8e-05
0.0e+
00
1.5e+
04
4.9e-02
9.7e+
02
50
1.7e+
07
3.2e+
02
1.5e-01
0.0e+
00
4.8e+
04
3.7e+
01
2.4e+
03
10
1.6e+
01
1.1e-01
1.2e-01
2.6e+
01
1.7e+
00
1.0e-01
1.1e+
00
Griewank
30
1.9e+
02
9.7e-01
2.1e-02
7.6e+
01
5.9e+
00
7.9e-01
1.8e+
00
50
3.7e+
02
1.3e+
00
2.6e-01
1.3e+
02
7.9e+
00
1.7e+
00
1.7e+
00
10
1.5e+
03
2.6e-06
1.4e-14
0.0e+
00
1.1e+
02
6.9e-11
1.0e+
01
Sphere
30
1.8e+
04
1.6e+
00
2.3e-03
0.0e+
00
6.9e+
02
1.8e+
00
8.3e+
01
50
3.6e+
04
2.8e+
01
1.7e-01
0.0e+
00
9.4e+
02
7.0e+
01
8.9e+
01
10
7.1e+
01
8.0e+
00
9.0e+
00
0.0e+
00
3.1e+
01
1.6e+
01
2.8e+
01
Rastrigin
30
3.0e+
02
5.1e+
01
3.5e+
01
0.0e+
00
1.8e+
02
9.5e+
01
1.9e+
02
50
5.3e+
02
1.1e+
02
6.5e+
01
0.0e+
00
3.1e+
02
1.8e+
02
3.3e+
02
10
2.8e+
05
6.8e+
00
7.1e+
00
9.0e+
00
4.0e+
03
6.0e+
00
4.4e+
02
Rosenbrock
30
2.1e+
07
2.9e+
02
1.0e+
02
2.9e+
01
6.3e+
04
2.5e+
02
1.1e+
04
50
5.2e+
07
3.4e+
03
2.5e+
02
4.9e+
01
9.9e+
04
1.5e+
03
1.4e+
04
1240011-17
![Page 18: P44_2012_IJAIT](https://reader035.fdocuments.in/reader035/viewer/2022070420/5695d0001a28ab9b0290810d/html5/thumbnails/18.jpg)
1st ReadingMay 4, 2012 16:6 WSPC/INSTRUCTION FILE S0218213012400118
J. L. Fernandez-Martınez & E. Garcıa-Gonzalo
! "! #!! #"! $!!!%"
#
#%"
$
$%"
&'()*+,
*-,./-*0+
(01#!23
,4*/+2,..0.55
6
6
! "! #!! #"! $!!!
$
7
8
9
#!4,6:0+1 ;7
*-,./-*0+
(01#!23
,4*/+2,..0.55
6
6
! "! #!! #"! $!! #
!
#
$
&
7<.*,=/+>
*-,./-*0+
(01#!23
,4*/+2,..0.55
6
6
! "! #!! #"! $!!#
$
&
7
"
8?)@,.,
*-,./-*0+
(01#!23
,4*/+2,..0.55
6
6
! "! #!! #"! $!!!%"
#
#%"
$
$%"
&A/B-.*1*+
*-,./-*0+
(01#!23
,4*/+2,..0.55
6
6
! "! #!! #"! $!!!
$
7
8
9
#!A0B,+C.0D>
*-,./-*0+
(01#!23
,4*/+2,..0.55
6
6
E?F
AA E?F
EE E?F
GG E?F
GE E?F
! "! #!! #"! $!!!%"
#
#%"
$
$%"
&'()*+,
*-,./-*0+
(01#!23
,4*/+2,..0.55
6
6
! "! #!! #"! $!!!
$
7
8
9
#!4,6:0+1 ;7
*-,./-*0+
(01#!23
,4*/+2,..0.55
6
6
! "! #!! #"! $!! #
!
#
$
&
7<.*,=/+>
*-,./-*0+
(01#!23
,4*/+2,..0.55
6
6
! "! #!! #"! $!!#
$
&
7
"
8?)@,.,
*-,./-*0+
(01#!23
,4*/+2,..0.55
6
6
! "! #!! #"! $!!!%"
#
#%"
$
$%"
&A/B-.*1*+
*-,./-*0+
(01#!23
,4*/+2,..0.55
6
6
! "! #!! #"! $!!!
$
7
8
9
#!A0B,+C.0D>
*-,./-*0+
(01#!23
,4*/+2,..0.55
6
6
E?F
AA E?F
EE E?F
GG E?F
GE E?F
! "! #!! #"! $!!!%"
#
#%"
$
$%"
&'()*+,
*-,./-*0+
(01#!23
,4*/+2,..0.55
6
6
! "! #!! #"! $!!!
$
7
8
9
#!4,6:0+1 ;7
*-,./-*0+
(01#!23
,4*/+2,..0.55
6
6
! "! #!! #"! $!! #
!
#
$
&
7<.*,=/+>
*-,./-*0+
(01#!23
,4*/+2,..0.55
6
6
! "! #!! #"! $!!#
$
&
7
"
8?)@,.,
*-,./-*0+
(01#!23
,4*/+2,..0.55
6
6
! "! #!! #"! $!!!%"
#
#%"
$
$%"
&A/B-.*1*+
*-,./-*0+
(01#!23
,4*/+2,..0.55
6
6
! "! #!! #"! $!!!
$
7
8
9
#!A0B,+C.0D>
*-,./-*0+
(01#!23
,4*/+2,..0.55
6
6
E?F
AA E?F
EE E?F
GG E?F
GE E?F
Fig. 10. Median convergence curves for different benchmark functions in 50 dimension in P2.
but it improves dramatically when we increase its number of iterations to several
thousands. Nevertheless, to increase the number of functions evaluations is not
always possible when applied to real problems due to the very high computational
cost.
This analysis done for point P2 has been also generalized for a grid of (ω, φ)
points within the first order stability region (figure 8).
1240011-18
![Page 19: P44_2012_IJAIT](https://reader035.fdocuments.in/reader035/viewer/2022070420/5695d0001a28ab9b0290810d/html5/thumbnails/19.jpg)
1st ReadingMay 4, 2012 16:6 WSPC/INSTRUCTION FILE S0218213012400118
Analysis of Two Novel Algorithms of the PSO Family: PP-GPSO and RR-GPSO
6. Conclusions
In this paper we present two more different members of the PSO family: PP-PSO
and RR-PSO. Both versions are deduced from the PSO continuous model adopting
respectively a progressive and a regressive discretization in velocities and accelera-
tions. Although they are PSO-like versions, PP-PSO has the same velocity update
than GPSO, and RR-PSO has the form of a PSO with constriction factor, their
behavior is very different from the PSO case. RR-PSO best parameters are concen-
trated along a straight line located in the complex zone of the first order stability
region, but are not in direct relation with the upper limit of the second order sta-
bility zone. Its behavior is very different from others family members including
PP-PSO. Finally with respect to the convergence rate, RR-PSO seems to provide
similar or even better results than PSO and PP-PSO seems to be most explorative
(higher median misfits). Knowledge of these two new versions can be very impor-
tant in the solution of very different optimization and inverse problems in science
and technology as shown in some applications in oil and gas and protein secondary
structure prediction.
Acknowledgments
This work benefited from one-year sabbatical grant (2008-2009) at the University
of California Berkeley (Department of Civil and Environmental Engineering) given
by the University of Oviedo (Spain), and by the “Secretarıa de Estado de Univer-
sidades y de Investigacion”of the Spanish Ministry of Science and Innovation. We
also acknowledge the financial support for 2009-2010 coming from the University
of California Berkeley, the Lawrence Berkeley National Laboratory (Earth Science
Division) and the Energy Resources Engineering Department of Stanford University
(Stanford Center for Reservoir Forecasting and Smart Field Consortia).
References
1. M. Clerc and J. Kennedy, The particle swarm—explosion, stability, and convergence ina multidimensional complex space, IEEE Trans. on Evolutionary Computation Febru-ary, 6 (2002) 58–73, doi:10.1109/4235.985692.
2. J.L. Fernandez-Martınez and E. Garcıa-Gonzalo, Theoretical analysis of particleswarm trajectories through a mechanical analogy, Int. J. of Comp. Int. Res. 4 (2008)93–104.
3. J.L. Fernandez-Martınez and E. Garcıa-Gonzalo, Stochastic stability analysis of thelinear continuous and discrete PSO models, IEEE Transactions on Evolutionary
Computation 15 (2011) 405–423, doi:10.1109/TEVC.2010.2053935.4. R. Poli, Dynamics and stability of the sampling distribution of particle swarm op-
timisers via moment analysis, Journal of Artificial Evolution and Applications 2008
(2008) 1–10, doi:10.1155/2008/761459.5. J.L. Fernandez-Martınez and E. Garcıa-Gonzalo, The generalized PSO: A new door
for PSO evolution, Journal of Artificial Evolution and Applications 2008 (2008) 1–15,doi:10.1007/s11721-009-0034-8.
1240011-19
![Page 20: P44_2012_IJAIT](https://reader035.fdocuments.in/reader035/viewer/2022070420/5695d0001a28ab9b0290810d/html5/thumbnails/20.jpg)
1st ReadingMay 4, 2012 16:6 WSPC/INSTRUCTION FILE S0218213012400118
J. L. Fernandez-Martınez & E. Garcıa-Gonzalo
6. J. L. Fernandez-Martınez and E. Garcıa-Gonzalo, The PSO family: Deduction,stochastic analysis and comparison, Swarm Intelligence 3 (2009) 245–273, doi:10.1007/s11721-009-0034-8.
7. Y.-L. Zheng, L.-H. Ma, L.-Y. Zhang and J.-X. Qian, On the convergence analysis andparameter selection in particle swarm optimisation,Proc. of the 2nd International Con-
ference on Machine Learning and Cybernetics (ICMLC’03), Xi’an, China, November,3 (2003) 1802–1807, doi:10.1109/ICMLC.2003.1259789
8. J. L. Fernandez-Martınez, E. Garcıa-Gonzalo, J. P. F. Alvarez, H. A. Kuzma and C. O.Menendez-Perez, PSO: A powerful algorithm to solve geophysical inverse problems.Application to a 1D-DC resistivity case, Jounal of Applied Geophysics 71 (2010) 13–25, doi:10.1016/j.jappgeo.2010.02.001.
9. J. L. Fernandez-Martınez and E. Garcıa-Gonzalo, Particle swarm optimization appliedto the solving and appraisal of the streaming potential inverse problem, Geophysics
75 (2010) WA3–WA15, doi:10.1190/1.3460842.10. A. Suman, J. L. Fernandez-Martınez, and T. Mukerji, (2011) Joint inversion of time-
lapse seismic and production data for Norne field, SEG Technical Program Expanded
Abstracts 30 (2011) 4102–4108, doi:10.1190/1.3628063.11. J. L. Fernandez-Martınez, E. Garcıa-Gonzalo, S. Saraswathi, R. Jernigan and
A. Kloczkowski, Particle swarm optimization: A powerful family of stochastic op-timizers. Analysis, design and application to inverse modelling, in Y. Tan, Y. Shi,Y. Chai and G. Wang, eds., Advances in Swarm Intelligence (Springer Berlin, Heidel-berg, 2011), pp. 1–8, doi:10.1007/978-3-642-21515-5 1.
12. S. Saraswathi and J. L. Fernandez-Martınez, A. Kolinski, R. Jernigan andA. Kloczkowski, Fast learning optimized prediction methodology (FLOPRED) forprotein secondary structure prediction, Journal of Molecular Modeling, accepted forpublication.
13. C. Trelea, The particle swarm optimization algorithm: Convergence analysis and pa-rameter selection, Information Processing Letters 85 (2003) 317–325, doi:10.1016/S0020-0190(02)00447-7.
14. PSO (2007), Standard PSO, http://www.particleswarm.info/standard pso 2007.zip15. B. Birge, PSOt — A particle swarm optimization toolbox for use with Matlab, Proc. of
the IEEE Swarm Intell. Symp (SIS 03) (2003) 182–186, doi:10.1109/SIS.2003.1202265.
1240011-20