27.06.2013 Views

Evolution and Optimum Seeking

Evolution and Optimum Seeking

Evolution and Optimum Seeking

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

Numerical Comparison of Strategies 211<br />

is in uenced by r<strong>and</strong>om numbers, the same rule was followed: namely, out of three tests<br />

the one with the best end result was accepted. In contrast to the strategy of Box, however,<br />

the evolution methods prove to be less dependent on the actual sequence of r<strong>and</strong>om<br />

numbers. This is especially true of the multimembered versions. Recombination almost<br />

always improves the chance of getting very close to the desired solutions. Fatal errors<br />

due to exceeding the maximum number range or dividing by zero do not occur by virtue<br />

of the simple computational operations in these strategies. Discontinuities in the partial<br />

derivatives, saddle points, <strong>and</strong> the like have noobvious adverse e ects. The search does,<br />

however, become rather time consuming when the minimum is reached via a long, narrow<br />

valley. The step lengths or variances that are set in this case are very small <strong>and</strong> impose<br />

slow convergence in comparison to methods that can perform a line search along the<br />

valley. The average rate of progress of an evolution strategy is not, however, a ected by<br />

bends in the valley, which would retard a one dimensional minimization procedure. Line<br />

searches only a ord a signi cant advantage to the rate of progress if there are directions in<br />

the space along which successful steps can be made of a size that is large compared to the<br />

local radius of curvature of the objective function contour surface. Examples are provided<br />

by Problems 2.14, 2.15, <strong>and</strong> 2.28. In these cases, long before reaching the minimum the<br />

optimal variances of the evolution methods have reached the lower limit as determined<br />

by the machine accuracy. The desired solution cannot therefore be approximated to the<br />

required accuracy. In Problems 2.14 <strong>and</strong> 2.15 the computation time limit did not allow<br />

the convergence criterion to be satis ed although it was actually progressing slowly but<br />

surely, the search was terminated.<br />

Di culties with the termination rule based on function values only occurred in the<br />

solution of one type of problem (Problems 2.11, 2.12) using the (10 , 100) evolution strategy<br />

with recombination. The multimembered method selects the 10 best individuals of a<br />

generation only from the current 100 descendants. Their 10 parents are not included in<br />

the selection process, for reasons associated with the step length adaptation. In general,<br />

the objective function value of the best descendant is closer to the solution than that<br />

of the best parent. In the case of the two problems referred to above, this is initially<br />

the case. As the solution is approached, however, it happens more <strong>and</strong> more frequently<br />

that the best value occurring in a generation is lost again. This is related to the fact<br />

that because of rounding errors in evaluating values near the minimum, the objective<br />

function behaves practically stochastically. Thus the population w<strong>and</strong>ers around in the<br />

neighborhood of the (quasi-singular) optimal solution without being able to satisfy the<br />

convergence criterion. These di culties do not beset the other search methods, including<br />

the multimembered evolution without recombination, because they do not come nearly so<br />

close to the optimum. The fact that the third problem of the same type (Problem 2.10)<br />

is solved without di culties in a nite time, even with recombination, can be considered<br />

a uke. Here too the minimum was reached long before the termination criterion was<br />

satis ed. On the whole, the multimembered evolution strategy with recombination is the<br />

surest <strong>and</strong> safest of all the search methods tested. In only 5 out of 28 cases is the solution<br />

not located exactly, <strong>and</strong> the greatest deviations of the variables were in the accuracy class<br />

=10 ;4 .

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!