Instead of manual tuning of algorithm's parameters, it is recommended to utilize automatic algorithm configuration software. Mostly beacuse it was shown that they increase manyfold the algorithm's perfomance. However, there are some differences among the proposed configuration software and beside those listed in (Eiben, Smit, 2011) it is important to gather experiances from the researchers. I would like to hear how does one decide on the stopping criteria, or values for parameters, for heuristic steps within the stochastic algorithm... there are so many questions.
As you mentioned, parameter tuning studies for a metaheuristic is quite important. Researchers should determine proper control parameters for their optimization problem to increase the success of the algorithm. However, many researchers uses algorithm parameters suggested by their developers as this is can be a time consuming task via a trial and error approach. Also, I agree that self-adaptive versions of these algorithms can increase both effectiveness and performance compared to their original versions. However, they can require definition of extra parameters as well in the algorithm. In my cases, I prefer to use original versions of the algorithms via a parameter tuning study. Besides, I use two termination criteria including a predefined maksimun generation number and a tolerans value. If the algorithm provides a misfit value less than the tolerans, it stops before the reaching maksimum number of generation. Sometimes I take into account a number of successive generations. For instance, if the solution do not improve during the last 30 generations, I stop the algorithm. This provides relatively decrease the high computation cost due to much execution of the forward equation. This is the biggest drawback of the global optimization compared to derivative-based approaches considering high-dimensional optimization problems.
Çağlayan Balkaya thank you for your answer. I also like the tolerance value when I want to compare two algorithms with cumulative empirical distribution graphics. Do you repeat executions for different values of seed? How much times do you repeat? I absolutely agree that no-improvement number of iterations helps to decrease computational costs. I found myself kicking my head to find an appropriate budget... you know - not to much, not to little. But it is driving me insane sometimes.
Tatjana Jakšić Krüger I run the algorithm consecutively at least 20-30 times to obtain statistical performance analysis. In my studies, generation numbers vary depending on both the algorithm used and the optimization problem considered. Do you mean getting different seeds in each independent run so as not to duplicate the results of the previous execution?
- Badges
- Science topic
- Similar topics
- Linguistics
- Composition Studies
- Writing