xiðk þ 1Þ¼ —ða — x — 1Þxi ðkÞ— xxi ðk — 1Þþ a1xl ðkÞþ a2xg ðkÞ ð8Þ
with xi ð0Þ ¼ x0 and xi ð1Þ ¼ ð1 — aÞx0 þ xv0 þ a1xl ð0Þþ a2xg ð0Þ。
Let xi ðk þ 1Þ ¼ xi ðkÞþ vi ðk þ 1Þ and xi ðk — 1Þ ¼ xi ðkÞ— vi ðkÞ。 Consequently, the widely used form of the PSO algorithm pre- sented by (1) can be derived in terms of (8)。
Consider another case of Dt ¼ Dk > 0 and t ¼ k。 We obtain the GPSO algorithm [13] as
x ðk þ DkÞ¼ c x ðkÞþ c x ðk — DkÞþ Dk2ða x ðkÞþ a x ðkÞÞ ð9Þ
with
c1 ¼ 2 — ð1 — xÞDk — aDk c2 ¼ ð1 — xÞDk — 1
Similarly, we rewrite the Eq。 (9) based on the position and the velocity as
( x ðkÞþa xg ðkÞ 。
vi ðk þ DkÞ¼ ð1 — ð1 — xÞDkÞvi ðkÞþ aDk。a1 l
xi ðk þ DkÞ¼ xi ðkÞþ vi ðk þ DkÞDk
2
a1 þa2
— xiðkÞ
ð10Þ
As pointed out by Fernández Martínez et al。 [13], the particle swarm movement controlled by the GPSO algorithm be- comes more elastic and less damped when Dk ! 0。
Remark 3。 Both the PSO algorithm (1) and the GPSO algorithm (10) can be derived by the same continuous-time model (6)。 The motivation of the GPSO algorithm proposed by Fernández Martínez et al。 [13] is that the search performance of the GPSO algorithm approaches to that of the corresponding continuous-time model when Dk ! 0。
3。2。Analysis on the continuous-time model of particle swarm optimization
In this subsection, we will analyze the convergence characteristic of the PSO algorithm。 Let ni ðtÞ ¼ xi ðtÞ— pi ðtÞ, where
a1 xl ðtÞþa2 xg ðtÞ
pi ðtÞ ¼
a1 þa2 , and introduce ni ðtÞ into the continuous-time model of the PSO algorithm (6)。 We obtain
€ni ðtÞþ p€iðtÞþ ð1 — xÞðn_ i ðtÞþ p_ i ðtÞÞ þ aðni ðtÞþ pi ðtÞÞ ¼ a1xl ðtÞþ a2xg ðtÞ
Then, we have
€ni ðtÞþ p€iðtÞþ ð1 — xÞðn_ i ðtÞþ p_ i ðtÞÞ þ aðni ðtÞþ pi ðtÞÞ ¼ api ðtÞ
Finally, we derive
€ni ðtÞþ ð1 — xÞn_ i ðtÞþ ani ðtÞ¼ —p€i ðtÞ— ð1 — xÞp_ i ðtÞ ð11Þ
Let y1 ðtÞ ¼ ni ðtÞ and y2ðtÞ ¼ n_ i ðtÞ。 Then, the Eq。 (11) can be rewritten as
。 y_ 1ðtÞ¼ y2ðtÞ