Hi, Dear Shervin Ataeian,

Applyed to different kind of problems, meta-heuristics don't have the same performance, some can be good in a domain and perform worst in others.

This is du to the adaptation of their exploration/exploitation mechanisms to the tackled problem.

Sincerely, Djamel.

"Learn, Teach, Respect and Think."

Djamel Zeghida Thank you for your answer. All of these algorithms are applied to the same problem. Otherwise, their performance cannot be compared properly.

Indeed, some meta-heuristics will and should be better than others, in this case, and that is exactly the point of the comparison!

What to expect from a comparison if not different performance of the applied approaches !!!!

Sincerely, Djamel.

"Learn, Teach, Respect and Think."

Djamel Zeghida Thank you so much for your answer. I just wanted to make sure that my comparison is valid. I knew that GA is slow compared to many other metaheuristic algorithms but in this case it converged to a solution that is not even close to the solution found by other metaheuristic algorithms and I didn't expect this. Therefore, I wanted to make sure that I do not have to run it for more iterations and I can give up on it.

You are welcome.

Kindly note that there are many different instantiations of "a GA." Depending on the sort of implementation of your GA it may include certain features that work well for one problem and not so well for another. Therefore, unless the specific ingredients are specified in detail one cannot expect a specific valid statement regarding a certain method.

Hello Shervin!

One important thing in a optimizatiom algorithms to find some good solution is maintain diversity of population. Is very common during process of algorithm occurs the elimination of diversity of individuals of population. One way to solve this issue is implement some function that promote a big variability of individuals. You can find this functionality in hibridy algorithms like GASA that, is a join Genetic Algorithm with Simulated Annealing algorithm.

because the behaviors of metaheuristic algorithms are different in any problem.

Hi

What is your problem and what other metaheuristics are you comparing with ?

Because GA is performant in discrete problems.

If your problem is continuous, it might not be relevant to GA, and according to the coding system of real numbers, it might not provide good results.

Peio Loubiere Hi, thank you for your answer, yes my question is discrete.

Even if GA is not giving good solution, you can simply report it. This will simply be considered as a negative result, i.e. a result that is against you hypothesis. Reporting this solution will also help the new researchers to know that GA is unable to find good approximate to global optimum solution for the problem of concern.

I would also like to suggest that GA or any other metaheuristic algorithm in its original form usually do not guarantee convergence to global best. They sometimes or many a times need some tuning of parameters. The issue you are facing maybe related to premature convergence to local bests. You can add some chaotic maps in original GA update equations. You can also work with different neighborhood topologies like Von Neuman. You can make a hybrid of GA with some other metaheuristic algorithms as well.

You can also look at the tuning parameter that deals with exploration and exploitation, and try to make it dynamic, maybe linearly, or with some chaotic map, or by using Levy Flights etc. Hopefully it ll work. If not, still not an issue. As No free lunch theorem states that if an algorithm performs good on problem A, it does not guarantee that it ll work good for problem B also.

Best of luck for your research.

This is strange..., what other algorithms are involved and what is your problem ?

Are you using a library version of GA or is it your own ?

Did you check the convergence of you global best solution accross all iterations ?

Did you test it on classical problems such as TSP or Knapsack whith known optima ?

One possible explanation is that standard GAs do not include local search as part of their implementation. Metaheuristics such as Iterated Local Search incorporate local search and that makes them very effective. One solution is to hybridize the GA with local search, which is sometimes called Memetic Algorithms.

Another explanation is that sometimes Crossover operators are not very effective. So a GA with a very effective crossover can have good performance, but another GA with a mediocre crossover operator can have poor performance.

I would recommend tackling the problem in two steps. The first thing you have to recognize is that metaheuristic algorithms (GA included) are stochastic optimization methodologies and therefore are prone to their initial algorithmic parameters. They have a sense of chaos to them in that different sets of parametric values yield notable variations in algorithmic results, stability and speeds. Ergo, one cannot claim to properly compare various metaheuristic methods without first properly adjusting each tuning parameter. There are established methodologies to perform "Parameter Sensitivity Analysis" for metahueristic techniques. Judging by how you seemingly change parameter values like "mutation rate" and "crossover rate" randomly, you can't hope to produce any satisfactory results. The best parameter sensitivity analysis method that I use is the " One-at-a-time (OAT)" method. For all practical purposes, its simple and efficient. It typically involves changing one of the parametric value at a time while keeping others at a nominal baseline value. You do these for the rest of the variables and note how each parameter effects algorithmic robustness. You can read more about the method on wikipedia: https://en.wikipedia.org/wiki/Sensitivity_analysis

After you have "properly" selected your parameters and you some how have still not produced good results in comparison with other metahueristic algorithms, you have no choice but to report your findings. Personally, you shouldn't be surprised if you produce subpar results with GAs. It's considered one of the older metahueristic algorithms that are out there and is known to have weak exploitative capabilities. I know this as a fact because I have investigated its potential feasibility in structural optimization problems as a masters thesis.

Finally, as a last piece of advice, in papers that focus on comparing algorithmic techniques, the method you go about doing that is of utmost importance. I would recommend reading this paper and attempt to adopt its comparison methodology: https://www.sciencedirect.com/science/article/abs/pii/S2210650219304080

Good luck

Not all GAs or metaheuristics perform the same for a given optimisation problem, so what you have observed in your experiments is not unusual I would say. Simply changing or trying different parameter settings is one of the easiest things to do for comparison purposes, but only doing this may not be enough to understand why the other metaheuristics obtained better results than your specific GA. An alternative is to carry fitness landscape analysis based on the search operators and solution representation you are using. This can give you deeper insights into how those metaheuristics are exploring/exploiting the landscape, but such analyses can be harder to do depending on which problems you are tackling.

Dear Shervin,

Firstly, you can report these poor results of GA with its control parameters without any problem, because your goal is not improving the performance of GA. In most of the optimisation problems, the convergence speed of GA is lower than other popular metaheuristics such as PSO, GWO and so on. Just to confirm, that would be great to double-check on the GA code to be sure the implementation of mutation, crossover and selection operators are correct.