Dear Milos Simic,

It 's an interesting question. In term of your question, Yes. If I want to compare the performance of algorithms in a fair manner, I will use the same random seed based on time.I use MATLAB software with this command (rand('state', sum(100*clock))).

Best Regards,

Soheila

Using the same set of seeds may make sense when you use the same metaheuristic. For example, when you want to compare distinct variants of the same method or when you want to fine-tune the parameters of the method for the problem.at hand

On the other hand, when comparing two different metaheuristics, it will not make much difference whether you use the same set of seeds or not.  The exception, perhaps, is when you want to make sure that both methods start with exactly the same initial conditions Other than that, the difference in the internal working of the methods will outweigh the difference in the sequence of random numbers 

Thank you for your answers.

Yujiang, I will read Your article, it seems interesting.

As for comparing metaheuristics and tuning the parameters, I do agree with H.E. Lehtihet. However, the question is how to choose the seeds. For example, in C we usually call srand(time(NULL)) to set the seeds to the number of seconds since Jan 1, 1970. This method works for most use cases as it ensures that we will always use different seeds (provided that we do not run the program more than once in the same second). However, if we decide to choose the seeds manually, then there is a risk that some of them will result in correlated sequences of random numbers, which will mean that results of the metaheuristic in question for those seeds will not be independent of one another.

Most implementations of C's rand() use LGC: https://en.wikipedia.org/wiki/Linear_congruential_generator . The same page contains parameters of LGC for different languages.

In fact, if your random number generator is a good one, any seed will be OK. But most of the random number generators that came with compilers are only very basic LCG generators.

If you are interest, you can refer to works by Pierre L'Ecuyer's. There are plenty of references for good random number generators.

An important component of the seed for generating random numbers is constituted by the current data, hours, mute and seconds. In any moment you will generate different numbers. it is very important in the most of applications. Ar least for me it has being very useeful

Best wishes, José Arzola

Just use a good quality random number generator.  Otherwise you will be testing, in a rather sophisticated way, the random number generator itself, not the properties of your metaheuristic algorithms.

In experimental research I use Mersenne Twister for generating random numbers. It is much better than algorithm implemented in old C with functions rand() and srand(). You can obtain source code directly from the authors that algorithm or find it somewhere on the Internet. It is available for plain C, C++ and other languages. Mersenne Twister is now also available in C++11 standard, and maybe (not sure) in some new C standard.

In every execution of the algorithm I use new seed, but along with result I always record which seed was used in that particular run of the algorithm. This is useful for, debugging, analyzing some unusual results etc.

No metaheuristic is worth its salt if its performance varies significantly with the choice of seed or even the method (of a good standard) of generating the random number sequences of a particular statistical distribution (pdf) from a given seed. Why? Because the power of a metaheuristic is in searching and speed of convergence, not on hitting the surface of the problem geometry in a particular manner. Of course, distribution of random numbers (e.g. normal, uniform, Levy, Gamma, etc) will matter.

I agree with Dr. Mishra statement: "No metaheuristic is worth its salt if its performance varies significantly with the choice of seed or even the method (of a good standard) of generating the random number sequences of a particular statistical distribution (pdf) from a given seed".

Anyway, if you want to use the same set of seeds for all the runs of different algorithms, you can generate (using a good random number generator) as many seeds as you want, store them in a file and pick the seed from there.

I do not consider seeds value as a relevant factor that may influence the behaviour of a metaheuristic. Implementation issues, parameter settings, test instances, etc are more crucial.

I agree with Dr. Pelta ... "I do not consider seeds value as a relevant factor that may influence the behaviour of a metaheuristic. Implementation issues, parameter settings, test instances, etc are more crucial"

In my experience (TS, TA mainly), a good heuristic with appropriate parameters for the problem will perform well regardless of the seed.