Some works in the literature (Goldberg/Bridges, for instance) have demonstrated that standard Genetic Algorithms (GAs) usually cause the building blocks disruption of the solutions. On the other hand, some current papers have shown that the use of a GA special implementation may be a viable alternative in order to overcome this issue. In this sense, I would like to know what you think about that. Please, try answering my question in a short clear response if possible.
Thank you very much in advance!
Conference Paper Variable Selection for Multivariate Calibration in Chemometr...
The building blocks are not a critical point in genetic algorithms , this depends a lot of the problem is more used when you know patterns with which the solution must meet .
They can guide you to lead people to find solutions more quickly, but limit the search space exploration .
Dear Lauro Cássio:
The building blocks problem depends on the nature of the concrete task you are solving. But there are a variety of algorithms that makes evolve solutions codes populations (one or more) helped by a diversity of operators, GA could be included in this variety. Some of these algorithms could be more effective for a diversity of tasks.
The first article I wrotte devoted to this method "The Integration of Variables method". an additional information you can find in other works that you can download in my researchgate page, like Analysis and Synthesis of Engineering Systems.
Excepting the GA, the rest of the algorithms included in the method doesn't present the building blocks problem.
Best wishes,
José Arzola
The disruption are necessary to any problems when a solution is searched using any form of metaheuristics. Without disruption your search is likely to find a local optima. Your metaheuristics needs to recover from the disruption, it is a balancing act.
Building blocks are a key issue for understanding the inner working of the crossover operator. If you want to know if a crossover is working properly, you can run the "headless chicken test" (don't remember who defined it). The idea is simple: run your algorithm in its current form. Then make the crossover operation as follows: take one individual X from the population, randomly generate an individual Y and cross X, Y. Inject the child in the population. If both algorithms achieve the same performance, then crossover definition is not able to exchange the building blocks. Potentially, the same results can be obtained just with mutation (when you cross an individual with a randomly generated one, it acts as a "macro" mutation"
It is not the answer you likely expected but some practical suggestions.
Using BBs you can search some sub-space of genotype space. And it is not only related to local optima. Sometimes it is necessary to avoid redundant call of fitness function for certainly insufficient solutions (some variables could be limited and restrictions could be strengthened while algorithm runs). It is also the way to organise parallel algorithm (to run several GA on subpopulations that created by certain BBs than analyse results and generate improved BBs).
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Control Theory
- Control Systems