Instead of starting with random guesses, is it beneficial to feed a know best into the population, while using genetic algorithm as the optimizing technique for a optimization problem. I would like to know how good is such an approach? What are the merits/demerits of such an approach?
Thanks in advance.
.
the following link explores this issue in a specific domain (TSP)
http://www.desarrollosostenible.deusto.es/media/publications/2014/conferencepaper/on-the-influence-of-using-initialization-functions-on-genetic-algorithms-solving-combinatorial-optimization-problems-a-first-study-on-the-tsp.pdf
to quote from their conclusion :
"As final conclusion of this research it can be highlighted the efficiency of using heuristic initialization functions. Anyway, the excessive use of them could decrease the exploration capacity of the GA, trapping the population in local optimums quickly. Therefore, the key is to maintain a balance between individuals initialized by functions, and individuals generated randomly."
.
I think this is what should be done, if only good chromosome is indeed known. This way a chance for premature convergence of the algorithm is diminished. It wouldn't be wise not to use the available information. Of course, one should not concentrate the initial population around this trial solution only, as it may be still near to the local, not global optimum ...
@ Marek, thanks for the update.
So you mean one should not use initial guesses.
.
the following link explores this issue in a specific domain (TSP)
http://www.desarrollosostenible.deusto.es/media/publications/2014/conferencepaper/on-the-influence-of-using-initialization-functions-on-genetic-algorithms-solving-combinatorial-optimization-problems-a-first-study-on-the-tsp.pdf
to quote from their conclusion :
"As final conclusion of this research it can be highlighted the efficiency of using heuristic initialization functions. Anyway, the excessive use of them could decrease the exploration capacity of the GA, trapping the population in local optimums quickly. Therefore, the key is to maintain a balance between individuals initialized by functions, and individuals generated randomly."
.
@ Fabrice, thanks for the update.
I am not using any function, but just one individual is initialized with a know best. That to only for 1st generation.Will that effect the final results.
Thanks
Rajesh.
I agree very much with the quote provided by @fabrice "the key is to maintain the balance ..." because even seemingly "bad" random individuals can give rise to much better fitted offsprings thanks to crossover and mutation.
If only one individual is initialized with a known best, I do not think it will affect much the final result as GA is a population-based optimization tool.
.
you might also benefit from the following (and references therein)
https://www.researchgate.net/publication/269040767_Seeding_the_Initial_Population_of_Multi-Objective_Evolutionary_Algorithms_A_Computational_Study
.
Article Seeding the Initial Population of Multi-Objective Evolutiona...
Yes u can do this, there will be less chance to converse at local min this problem can be handle by mutation process.
why u r not interested in random generation?
I think it will not affect GA much because with your conventional operators like cross-over and mutation next generation will replace the existing solution ( although previous population is responsible for next population), so with subsequent generation your initial population will might be completely replaced.
@ Lokesh, thanks for the info. I was eager to now the influence of initialization on GA.
I think if one has some idea about initializing the population by some fitness function, i.e. if the initial population is good to fit closely with the fitness function, then initialing the population will help to get the solution easily.
For ex. In my case, I want to use GA for content based image retrieval. The system will retrieve the relevant image which are similar to the provided query image using some feature vector calculation and then finding the Euclidean distance.
After retrieving the relevant images, the user will choose the images from the result which are similar and which are not similar. The images that user will choose as similar to the query image, I will use that images as my initial population for the genetic algorithm and hope that in next iteration the retrieved images would be much better than the first iteration.
- Badges
- Science topic
- Similar topics
- Bioinformatics
- Bioinformatics and Computational Biology