27.06.2013 Views

Evolution and Optimum Seeking

Evolution and Optimum Seeking

Evolution and Optimum Seeking

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

Numerical Comparison of Strategies 179<br />

S. A. Lill (1971) Note on Algorithm 46<br />

Z. Kovacs (1971) Note on Algorithm 46<br />

Some of the parameters a ecting the accuracy were altered, either because the small<br />

values de ned by the author could not be realized on the available computer or<br />

because the closest possible approach to the objective could not have been achieved<br />

with them.<br />

Simplex method of Nelder <strong>and</strong> Mead:<br />

R. O'Neill (1971) Algorithm AS 47, function minimization using a<br />

simplex procedure<br />

A complete program for the Rosenbrock strategy:<br />

M. Machura, A. Mulawa (1973) Algorithm 450, Rosenbrock function minimization<br />

This was not applied because it could only treat the unconstrained case.<br />

The same applies to the code for the complex method of M. J. Box:<br />

J. A. Richardson, J. L. Kuester<br />

(1973)<br />

Algorithm 454, the complex method for constrained<br />

optimization<br />

The part of the strategy that, when the starting point is not feasible seeks a basis<br />

in the feasible region, is not considered here.<br />

Whenever the procedures named were published in ALGOL they have been translated<br />

into FORTRAN. All the other optimization strategies not mentioned here have also been<br />

programmed in FORTRAN, with close reference to the original publications. If one wanted<br />

to repeat the test series today, amuch larger number of codes could be made use of from<br />

the book of More <strong>and</strong> Wright (1993).<br />

6.3.3 Results of the Tests<br />

6.3.3.1 First Test: Convergence Rates for a Quadratic Objective Function<br />

In the rst part of the numerical strategy comparison the theoretical predictions of convergence<br />

rates <strong>and</strong> Q-properties will be tested, or, where these are not available, experimental<br />

data will be supplied instead. For this purpose two quadratic objective functions are used<br />

(Appendix A, Sect. A.1). In the rst (Problem 1.1) the matrix of coe cients is diagonal<br />

with unit diagonal elements, i.e., a scalar matrix. This simplest of all quadratic problems<br />

is characterized by concentric contour lines or surfaces that can be represented or<br />

imagined as circles in the two parameter case, spheres in the three parameter case, <strong>and</strong><br />

surfaces of hyperspheres in the general case. The same pattern of contours but with arbitrary<br />

monotonic variation in the objective function occurs in the sphere model for which<br />

the average rates of progress of the evolution strategies could be determined theoretically<br />

(Rechenberg, 1973 <strong>and</strong> Chap. 5 of this book).<br />

The second objective function (Problem 1.2) has a matrix of coe cients with all nonzero<br />

elements. It represents a full quadratic problem (except for the missing linear term)

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

Saved successfully!

Ooh no, something went wrong!