12.01.2015 Views

Evolutionary Computation : A Unified Approach

Evolutionary Computation : A Unified Approach

Evolutionary Computation : A Unified Approach

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.

218 CHAPTER 7. ADVANCED EC TOPICS<br />

10<br />

8<br />

Fitness: Best in Population<br />

6<br />

4<br />

2<br />

ES<br />

GA<br />

0<br />

0 100 200 300 400 500 600 700 800 900 1000<br />

Trials<br />

Figure 7.4: EA response to more rapid oscillating change.<br />

of change. Increased diversity can be accomplished in a number of ways. The first and most<br />

obvious strategy is to weaken selection pressure. The difficulty with this approach is getting<br />

the right amount of selection pressure without aprioriknowledge about the dynamics of<br />

the environment. Too much pressure (typical of many static optimizers) results in poor<br />

recovery performance. Too little pressure produces overall mediocre performance.<br />

An alternative is to use some form of crowding (De Jong, 1975) or niching (Goldberg,<br />

1989). Both approaches allow for strong initial selection, but restrict considerably the ability<br />

for individuals to take over the population.<br />

An interesting alternative is to use some form of restricted mating and selection. Island<br />

models and diffusion models restrict selection and mating to local subpopulations, thus<br />

maintaining more diversity. Alternatively, the use of “tag bits” for mating restrictions<br />

permits more dynamically defined “species” (Spears, 1994).<br />

So far, we have focused on maintaining diversity. A different strategy appears to be more<br />

useful when the environment undergoes occasional abrupt changes. In such cases the cost<br />

of maintaining diversity can be quite high. Instead, one can focus on providing diversity<br />

“on demand”. The difficulty here is in recognizing when the need arises without apriori<br />

knowledge about the landscape dynamics. One approach is to monitor and detect significant<br />

environmental changes, and then trigger something like a hypermutation operator (Cobb<br />

and Grefenstette, 1993) to produce the needed diversity. Alternatively, self-tuning operators<br />

such as the “1/5” rule for adapting mutation step size (Bäck and Schwefel, 1993) appear to<br />

be useful.

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

Saved successfully!

Ooh no, something went wrong!