Dear Miguel,

GA simulates the survival of the fittest phenomenon which implies that best individuals are more likely to pass to next generations. If I understand you correctly, probably your problem is with the choice of the selection operator.

Thank you for your replies. To account for stochasticity I extract the average fitness from 5 runs for each combination of parameters. However it does occasionally happen that the best fitness is found half way through and does not necessarily reflect the mode of the parameter values in the last generation.

Maybe the same combination of parameters appears on next generations but never with the same fitness (due to the stochastic nature), so even if it was the overall best during the process, it was a one-time occurrence and doesn't reflect the "stability" achieved after n more generations. So selecting the overall best fitness achieved during the GA run is probably not a good solution here, but instead looking at the most frequent set of parameter values in the final generation, assuming that most of the less fit choices got eliminated in the process. This is my line of though at the moment, but I am not sure if it is legitimate and acceptable to do so.

Dear Miguel:

  the GA is "just" a tool to solve your problem. You can perform any analysis you want: even to construct a histogram of the values that each variable attain at the end of the run. Also note that different solutions may have the same fitness, so a close inspection of the individuals in the final population is needed.

Regarding obtaining the best solution in the middle of the run, this is not a problem. Tuning the GA parameters is not trivial. If a run is fast for you, then perhaps you can enlarge the population size.

The most important question to answer is: is the GA able to solve the problem?

Regards

David

Thank you for your replies.

Esra'a: yes, the algorithm is picking the "overall best" as the solution, which would be 3 in my example. I could blindly accept this, but sometimes I look at the final generation, where supposedly the fittest have survived, and the 5 is more common than the 3. It may sometimes be the case that both values provide similar fitness, as David said.

David: Ok, that is good to know. Because I'm new to this I didn't know if I should blindly accept what the algorithm tells me it's the "best", or if I should go for a more in depth analysis of the final "survivors". If there are any disparities I can probably look at both ways of picking the best and then critically analyse the details in the discussion. I understand that picking the GA parameters is an art form, and probably my picks are not the best, but I used this method to demonstrate that I can minimise the output of the model until it is below a certain threshold, so I am happy if that is achieved, even if it's a local optimum. Not a huge deal if the GA parameters are not the absolute best in this case.

Miguel:

another suggestion: you can collect the overall best solutions from several runs and then check them carefully. I suppose you are dealing with real coded solutions (your variables are real), so if you put the solutions one below the other, you end with a matrix.

Then, you can analyze this matrix "by columns" to observe which are the good values for each parameter.

You can send me a private email if you need further assistance (I would be interested in your problem)

Regards

David

It is possible that your problem has more than one (local) optimum, not just only a global one.  The algorithm may be occasionally trapped in one of local optima.

David: yeah, i don't have that many runs of the GA though... I sent you a more detailed description through private message.

Marek: yes, I am aware I may have found a local optimum, but in this case it does not matter a lot if the output is satisfactory (under a threshold). And because the outcome is variable, the same combination of inputs may be better or worse depending on the random seed, and that is why I use the average of 5 runs. At least I hope I end up with a better solution overall than simple guessing or trial and error...

Dear Miguel,

Since you're performing a single objective optimization and it aims to minimize a function, you should look for the best minimum (in your case the individual with the fitness of 3) through the evolution process. As you suggested these evolutionary algorithms are stochastic methods and should be run multiple times to provide robust results. the best individual through the evolution process during multiple runs is your final optimal solution.

Dear Miguel,

If you say that sometimes the best solution is found in an intermediate generation, and does not always survive to the last generation, it is because you are not using elitism properly. As far as I know, elitism has nothing to do with the parent selection mechanism: elitism just means that, at the end of the current generation, after selecting parents and crossing+mutating them to create a new population, the best individual from the previous population replaces one individual from the newly created population (most commonly, it replaces the worst individual of the new population, although the selection of the individual to be replaced can be done differently). This is done in every generation, hence the best solution found so far always survive for sure and will get through to the last generation.

Hope it helps

Yes elitism ( 2 to 3 %)  protect the best genes . We don't have people with the quality of  Buddha or Jesus or ancient yogis with us now though evolution guarantees improvement in DNA .May be the fitness is enhanced in a different dimension. Likewise in GA simulation also it happens .  

Elite preservence will help you out from searching the already produced good solution in the previous genrations. It is an existing approach used in the genetic algorithms

Hi

Try to set the bounds of each variable, sometimes it help to save the best solution. You can also review to cost function because this problem is generally related to the accuracy of fitness function.

There is no best solution.. Only optimum solution is searched.. 

By "best solution" he means the "best solution found during the search process", from the beginning to the end of the search. Of course, we search for the optimal solution but most of the times we will not find it because the problem is really hard. What the GA returns is just the "best solution the algorithm has been able to find", which most likely does not mean the true optimal solution. No Genetic Algorithm provides a theoretical guarantee of finding the true optimal. 

So there IS a "best solution found" concept.

U may consider 3 as best solution as 5 may not satisfy the constraints u have defined or 3 might be the best solution satisfying your constraints. If you still have doubt u may check this by running few more iterations. I recommend to run few more iterations i. Your case