多模态优化论文




range. In this context, two cases are considered. In the first one, a random particle is chosen for the local search, then the new particle replaces the old one if its fitness is better. In the second one, the new particle replaces the old one based in some probability, just like simulated annealing. In this work, we combine a new local search algorithm, proposed in the Cortes' work [8], with canonical PSG. This new local search algorithms is also a hybrid algorithm combining features of clonal selection [9], genetic mutation and hill climbing [10], enhancing the capacity of exploration and exploitation of the canonical PSG.
For this sake, the paper is organized as follows: Section II presents the canonical PSG, the used local search times the s individual, then the population undergo a nonuniform mutation from genetic algorithms and only the best solution goes to the next generation.
s {::: initialize_solution(n); Eval(S); while NOT (Termination Criteria) do P +-Clone(s, c);
pI +-Mutation(P); s +-Eval(pl);
end while
Fig. 1. Pseudocode for LHS Algorithm
algorithms and the proposed one (PSOLHS); Section III shows the benchmarks functions used in our experiments; Section IV discusses the computational experiments; finally, Section V presents the conclusions and future works.
II. THE ALGORITHMS
A. Particle Swarm Optimization
The algorithm consists of particles that are placed into the search space. Each particle moves combining some aspects of its own history position and the global position. All particles move to next position and probably the swarm moves towards the potential optimum, in the next iteration.
Xp A particle represents a position in the search space as = (xL xL ... , x vf) f)· Further, a particle has a velocity Vi D = (v}, v; , ... ,which is used to determine the new position of the particle in the next iteration, where D represents the problem dimension. The new position is determined by means of the Equations and 1 and 2, where W represents the inertia weight, Cl C2 are acceleration constants, rl and r2 are random number in the range [0,1], p1 is the best position of the particle, and gd is the global optima of the swarm so far.
(2)
The velocity in any dimension is narrowed by a maxi
mum velocity V max in the range [-V
max, V max]. When the velocity of a particle violates this range we update it using
the Equation 3. In other words, we chose a new random
velocity within the range [-V max, V max].
vf = -Vmax + (randO X (Vmax -(-Vmax))) (3) B. Local Hybrid Search -LHS
The Local Hybrid Search was proposed in Cortes'
work [8]. Basically, the LHS works according to the pseudocode presented in Figure 1, where s represents the first random solution, P is a population created by cloning c In the original work, the LHS has some variations, so we used a variation named HNURT (Hybrid Non-Uniform Mutation with Random T) that presented the best results. In this approach, the maximum number of elements to be muted are obtained by the Equation 4, where n is the number of genes/dimension, Gmax is the max generation number and t is the current generation.
x (G Gmax max
-t)
+ lJ)
(4)
Actually, the Equation 4 is a linear equation that pro
duces a number between 1 and n (only if Gmax :::: n) based on the current iteration. In other words, the number of elements undergo mutation (nm) is a random number belonging to [1, n] in the initial generations and equals 1 in the last one.
Further, a position is mutated using non-uniform mutation according to Equation 5, where T E {O, I} is randomly generated.
c
=
{
se T = ° se T = 1
(5)
The function is calculated by Equation 6, where r is a random number belonging to [0,1], b is a parameter chosen by user, t is the current generation and Gmax is the maximal generation number. This equation is similar to non-uniform mutation of genetic algorithms excepting by the constant 0.5.
(6) In this context, we can notice that non-uniform muta
tion combined with nm tries to explore the search space
in the early iterations whereas exploit it in the final ones. All in all, the probability of function returns a number
close to zero increases as the algorithm advance.
C. The Proposed Algorithm -PSOLHS
The proposed algorithm uses the LHS for enhancing the exploration and exploitation capabilities of PSO, applying the LHS in the best particle according to the pseudocode presented in Figure 2.
热门文档