多模态优化论文
TABLE I
BENCHMARKS FUNCTIONS
-Xi X sin( fo)
L: =1 cos xT -exp( L: =1 cos 27rxiJ 13(x) -20 exp(
+20 + e
4000
= = = =
Benchmark Functions Domain Min Name Separable
[-500,500]
[-5.12,5.12] [-32,32] [-600,600] [-50,50]
-12569.5 0 0 0 0
Schwefel's Rastrign's Ackley's Griewank's Penalized
yes yes no no no
-1)2[1 + 10sin2(7rYi+Il]
4) 100, 10, +(Yn -1)2} + L: =1 U(Xi,
Yi 1 + (Xi + 1)
k(Xi -a)=, Xi> a 0, -a < Xi < a u(xi,a,k,m)
[1 + sin2(37rxi+Il]+ (Xn -1)[1 + sin2(27rXn)]} + 5, 100, 4) U(Xi,
=
=
_
[-50,50] 0 Penalized no
TABLE II
PARAMETERS FOR PSO, PSOLSH
PSO/PSOH
100 #particles
0.72984
the last solution, improved by local search, and the final
one (Mean of PSOLHS) is not far-away.
TABLE III
SUCCESS RATE OF THE LOCAL SEARCH
F.
5
#It. 1800
1000 300 400 300 300
Rate 259.3 284.8 130.1 139.9 147 147.9
Last It. 461 300 222 194 240 261
Last Solution -12569.48662 2.68E-06 0.000940578 0.05143648 2.61E-08 2.56E-06
a local minima in h, 12 and is, and seems to have similar performance in 14, 15 and, specially, in 16. However, looking into the numbers in Tables IV and Table V we can notice that PSOLHS also outperformed the other algorithms in 15 and 16 and, on the other hand, we have to accept the null hypothesis in 14.
In 12 LHS seems to evolve faster than PSOLHS, however, PSOLHS ends up with better results after about 700 iteration. This behavior is inverted in is where PSOLHS evolves faster than LHS, probably because is is more complex and no separable. The same idea is repeated in 16 for the canonical PSO, but looking at Table IV we can observe that PSO is also trapped in a local minima while in PSOLHS the evolving process carried out.
Table III presents the success rate of the local search according to the number of iterations (#Iterat.). The column Rate represents the mean of the number of times that the local search in PSOLHS enhances the solution. The column Last It. presents in which iteration the enhanced of the solution happened for the last time. The last improvement of the solution by the local search is illustrated in the column Last Solution. The last column shows the mean accomplished by PSOLHS.
In this context, we can notice that the final solution was obtained practically in the iteration 461 in the function h, becoming the searching for better solutions a hard task. The exploration capability is clear in function 12. On the other hand, the LHS take the advantage of the exploitation capability in functions is, 14, 15 and 16 because the solutions were enhanced close to the final iteration (#It versus Last It.), and the differences between
h 12 13
f4
fs f6
Mean of PSOLHS -12569.48662 2.42043E-13 8.78284E-05 0,01547516 5,47554E-09 3,06636E-07
Table IV shows a closer comparison between the canonical PSO and PSOLHS, where the second column represents the number of iterations of each algorithms which were chosen based on Yao's work [14]. So, taking the ttest into account, the function 14 has the same behavior in both algorithms (accepting the null hypothesis), probably because the search space is quite huge and the function is not separable, consequently, becoming the function very hard to optimize. The PSOLHS clearly outperformed the results of the canonical PSO, in the other function.
TABLE IV
COMPARISON BETWEEN PSOLHS AND PSO
h 12 13
f4
#Iter. 9000 5000 1500 2000 1500 1500
PSO -6976.8551 (519,545068) 55.4608487 (15.2414901) 2.39334217 (1.33037127)
fs f6
(0.03460991) 1.07585499 (2.60960767) 1.88703848 (4.13262833 )
0.02670347
PSOLHS
-12569.4866 (5,2147E-07) 2.4204E-13 (3.9627E-13) 8.7828E-05 (6.2498E-05)
t
59.9340777 20.26002274 10.01605827
1.194001145
(0.02005646) 5.4755E-09 (1.6195E-08) 3.0664E-07 (7.7303E-07)
0.01812523
2.295405212 2.542349096
Table V presents a comparison between PSOLHS and pure LHS. Again, the PSOLHS algorithm overcame the
热门文档