16.01.2013 Views

An Introduction to Genetic Algorithms - Boente

An Introduction to Genetic Algorithms - Boente

An Introduction to Genetic Algorithms - Boente

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.

(4.8)<br />

(This is simply the definition of proportional selection.) Thus, given and F, we can easily find , and<br />

vice versa. Vose and Liepins presented most of their results in terms of .<br />

Given these preliminaries, Vose and Liepins's strategy is <strong>to</strong> define a single "opera<strong>to</strong>r" G such that applying G<br />

<strong>to</strong> will exactly mimic the expected effects of running the GA on the population at generation t <strong>to</strong> form the<br />

population at generation t + 1:<br />

(4.9)<br />

Then iterating G on will give an exact description of the expected behavior of the GA. (This is quite<br />

similar <strong>to</strong> the types of models developed in population genetics by Fisher and others; see, e.g., Ewens 1979.)<br />

To make this clearer, suppose that the GA is operating with selection alone (no crossover or mutation). Let<br />

E(x) denote the expectation of x. Then, si(t) is the probability that i will be selected at each selection step,<br />

Let ~ mean that and differ only by a scalar fac<strong>to</strong>r. Then, from equation 4.8, we have<br />

which means<br />

Chapter 4: Theoretical Foundations of <strong>Genetic</strong> <strong>Algorithms</strong><br />

This is the type of relation we want (i.e., of the form in equation 4.9), with G =F for this case of selection<br />

alone.<br />

These results give expectation values only; in any finite population, sampling errors will cause deviation from<br />

the expected values. In the limit of an infinite population, the expectation results are exact.<br />

Vose and Liepins included crossover and mutation in the model by defining G as the composition of the<br />

fitness matrix F and a "recombination opera<strong>to</strong>r" that mimics the effects of crossover and mutation. (Vose<br />

and Liepins use the term "recombination" <strong>to</strong> encompass both the crossover and mutation step. I will adopt this<br />

usage for the remainder of this section.) One way <strong>to</strong> define is <strong>to</strong> find ri,j(k),the probability that string k<br />

will be produced by a recombination event between string i and string j, given that i and j are selected <strong>to</strong> mate.<br />

If ri,j(k) were known, we could compute<br />

In words, this means that the expected proportion of string k in generation t + 1 is the probability that it will be<br />

produced by each given pair of parents, times those parents' probabilities of being selected, summed over all<br />

possible pairs of parents.<br />

105

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

Saved successfully!

Ooh no, something went wrong!