多模态优化论文
X+---initialize--Bwarm(n); V+---while initialize_velocities(n); NOT (Termination Criteria) do eval(X); s xl +-if +-selecLone_particle(X) Fitness(xl) HNURT(s,K)
< Fitness(s) then end p +-xl
Fitness(xl) if if < gbest then end gbest if +-xl V end X+---+-update_velocity(V); while update_positions(X); eval(X); if min(Fitness(X)) < gbest then end gbest if +-x Fig. 2. Pseudocode for PSOLHS
The proposed algorithm tries to enhance the fitness of the best particle using LHS algorithm before the swarm search, where K is the number of iterations that the LHS
can use for. We have been applied the local search to the
best particle because Chen [7] proved to be the best choice.
Afterwards, the algorithm executes the canonical PSG,
updating the position and the velocity of particles. Indeed,
this particular order provides a better initial solution to
the canonical PSG.
III. BENCHMARK FUNCTIONS
In this work, we have used some known multimodal
benchmarks function to test if the performance of an
algorithm is better than another. In all presented cases
the benchmark functions must be minimized.
Each function has basically two important features:
separability and multimodality. A function is multimodal
if it has two or more local optima. It is separable if it can be
rewritten as a sum of p functions of just one variable. Non
separable functions are more difficult to optimize because the accurate search direction depends on two or more elements of the solution vector. The problem is even more difficult if the function is also multimodal. The dimensionality of the search space is another important factor in the problem complexity, because the number of local optima increases with the problem dimension. A study of the dimensionality problem and its features was carried out by Friedman [12]. We have used 6 benchmarks functions as shown in Table I, where all of them are multimodals. The domain of each function, the global optima, its name and whether the function is separable or not is presented, as well. The Schwefel's function, also known as Surface Schwefel,
is composed of a great number of peaks and valleys. The function has a second best minimum far from the global minimum where many search algorithms are trapped. Furthermore, the global minimum is near the bounds of the domain.
The Rastrigin's function was constructed from Sphere (f(x) = 2:x2) and its contour is made up of a large number of local minima, whose value increases with the distance to the global minimum. Therefore, the algorithm
could be trapped if genes are produced far from the global optima but into domain. The Ackley's function has an exponential term that covers its surface with numerous local minima. The complexity of this function is moderated, thus the search strategy must combine the exploratory and exploitative components efficiently. The Griewank's has a product term that introduces interdependence among the variables. The aim is the failure of the techniques that optimize each variable independently. As in Ackley's function, the optima of Griewank's func
tion are regularly distributed. Finally, Penalized functions, also known as Generalized Penalized Functions, are similar to Ratrigin's function, nonetheless, penalized functions are
no separable. IV. EXPERIMENTAL RESULTS The experiments have been conducted in a Intel i5
processor 2.67Ghz, 4GB (RAM), 500GB (HD) and Matlab
2010. In order to compare the algorithms variations, we
solved each benchmarks through 31 independent trials and
dimension equals 30 (n =
30). This number of executions
is based on the central limit theorem, which claims that
with 30 or more runs a sample will present a normal distribution, allowing statistical inferences [13], like a ttest, for instance.
Considering a t-test, our null hypothesis is that there
are no differences between means, i.e., Ho : ILl = ILtest with level of significance 2, which
indicates a two tailed a =
0.5 with 30 degrees of freedom. Therefore, any result of t,
calculated by Equation 7, has to be in the range [-2.042,
2.042] for accepting the null hypothesis, where ILi are the
means being compared, O"i are the standard deviations and
n is the number of trials.
t =
"i (7) + ,,
Moreover, in order to make a fair n comparison between algorithms, the same number of functions evaluations have been done. For instance, the h function executes 9000 iterations in PSG and LHS in each trial, then the PSOLHS executes 1800 iterations times 5 for the local search (1800 using the parameters presented in x 5 = 9000). All experiments Table II, have been where conducted ub and lb are the upper bound and the lower bound of the search space, respectively.
The Figure 3 presents the performance of PSO, LHS and PSOLHS algorithms. Clearly, PSO has got trapped in
热门文档