There is a common agreement that the evolutionary method (EM) is a space searching method inspired by the evolution theory and genetics. Usually, engineers use the EM as an optimization tool to perform e.g. design or control tasks, rather than a tool for simulating the evolution. Therefore, it is useful to characterize the evolutionary method in terms of numerical optimization with much smaller stress on the biological metaphor.
EM consists in searching the space of candidate solutions by sampling it on random, according to a certain probability distribution, referred to as „sampling distrribution” (SD). When EM is treated as an optimization tool, a crucial issue is to characterize the dynamics of the SD and, the areas where the candidate solutions will be most frequently sampled. It is believed that such „most interesting points” should be the points where the objective function takes its maximum. In the presentation, we provide counterarguments to that hypothesis.
The role of genetic operators are lively discussed. We argue that the mutation is indispensable, whereas the crossover can slow down the dynamics of the EM. We also discuss the issue of the population diversity, which is believed to be the factor that has to be maintained. We evidence that the diversity level is a result of the selection process rather than a separate factor that has to be specially controlled.
When solving a real world problem, one has to provide a reasonable solution in an acceptable time. We argue that is it more adviceable to combine the evolutionary based search (which is robust to local maxima, but very imprecise) with local optimization methods (which in turn are rather precise and fast in finding local maxima). Possible methods for such hybridization are briefly discussed.
In case of practical problems, it is needed to introduce constraints to solutions infeasible solutions. EM allow for a variety constraint handling technques, which will be briefly mentioned and commented. |