11.07.2015 Views

View - Universidad de Almería

View - Universidad de Almería

View - Universidad de Almería

SHOW MORE
SHOW LESS
  • No tags were found...

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

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

128 Eligius M.T. Hendrix1.1 EffectivenessFocusing on effectiveness there are several targets a user of the algorithm may have:1. To discover all global minimum points. This of course can only be realised when thenumber of global minimum points is finite.2. To <strong>de</strong>tect at least one global optimal point.3. To find a solution with a function value as low as possible.4. To produce a uniform covering: This i<strong>de</strong>a as introduced by [3], can be relevant for populationbased algorithms.The first and second targets are typical satisfaction targets; was the search successful or not?What are good measures of success? In the ol<strong>de</strong>r literature, often convergence was used i.e.x k → x ∗ , where x ∗ is one of the minimum points. Alternatively one observes f(x k ) → f(x ∗ ).In tests and analyses, to make results comparable, one should be explicit in the <strong>de</strong>finitions ofsuccess. We need not only to specify ɛ and/or δ such that‖x k − x ∗ ‖ < ɛ and/or f(x k ) < f(x ∗ ) + δ (2)but we should also specify whether success means that there is an in<strong>de</strong>x K such that (2) is truefor all k > K. Alternatively, a record min k f(x k ) may have reached level f(x ∗ ) + δ and this canbe consi<strong>de</strong>red a success. Whether the algorithm is effective also <strong>de</strong>pends on the stochastic natureof it. When we are <strong>de</strong>aling with stochastic algorithms, the effectiveness can be expressedas the probability that a success has been reached. In analysis, this probability can be <strong>de</strong>rivedwhen having sufficient assumptions on the behaviour of the algorithm and in numerical experimentsit can be estimated by looking over repeated runs how many times the algorithmleads to convergence. We will give some examples of such analysis.1.2 EfficiencyGlobally efficiency is <strong>de</strong>fined as the effort the algorithm needs to be successful. A usual indicatorfor algorithms is the (expected) number of function evaluations necessary to reach theoptimum. This indicator <strong>de</strong>pends on many factors such as the shape of the test function andthe termination criteria used. The indicator more or less suggests that the calculation effortof function evaluations dominates the other calculation effort of the algorithm. Several otherindicators appear in literature that can be consi<strong>de</strong>red to be <strong>de</strong>rived from the main indicator.In nonlinear programming (e.g. [5] and [2]) the concept of convergence speed is common.It is a limiting concept on the convergence of the series x k . Let x 0 , x 1 , . . . , x k , . . . converge topoint x ∗ . The largest number α for which‖x k+1 − x ∗ ‖limk→∞ ‖x k − x ∗ ‖ α = β < ∞ (3)gives the or<strong>de</strong>r of convergence, whereas β is called the convergence factor. In this terminology,among others the following concepts appearlinear convergence means α = 1 and β < 1quadratic convergence means α = 2 and 0 < β < 1superlinear convergence: 1 < α < 2 and β < 1, i.e. β = 0 for α = 1.Mainly in <strong>de</strong>terministic GO algorithms, information on the past evaluations is stored in thecomputer memory. This requires efficient data handling for looking up necessary information

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

Saved successfully!

Ooh no, something went wrong!