1
$\begingroup$

I'm using the genetic algorithm to select the architecture of a neural network, however it is yet to provide good results.

I have an initial population size of 40 (each individual represents a possible architecture, made up of two genes, represented by integers), and I have evolved them for 10 generations.

The mean fitness of the individuals in the generation improves for the first 3 generations, then plateaus at a value slightly above the fitness of the best individual. The fitness of the best individual stays the same over all 10 generations. In maths terms, I think this is known as being stuck in a local minimum.

I find it very unlikely that I found the best solution in my first generation, so something seems to be stopping the individuals from evolving into anything better than this. Is there anything I can do to try and make the the fitness of the best individual improve after each generation?

$\endgroup$
1
$\begingroup$

Genetic algorithms are often used when the search space is high dimensional and non co-convex. Since you only have two dimensions you might be better off doing a grid search over a reasonable range of values.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ I've considered this. The two variables are also constrained between 1 and 25, meaning there is a total of 625 possible solutions. It would make sense to do an exhaustive search over the entire solution space, but the cost function used to evaluate each pair of variables takes on average about 5 minutes (less for smaller pairs, more for bigger pairs). I thought I'd use the GA to try and selectively sample the best points in this solution space, so that the time taken to search it is reduced. $\endgroup$ – Blue7 Jun 1 '14 at 15:56
  • $\begingroup$ You're not going to save all that much time considering your population size is 40 and you were doing 10 generations. So that's 400 vs 625 evaluations. $\endgroup$ – Aaron Jun 1 '14 at 18:04

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.