By combining "many" different heuristics approaches available in

"Journal of Heuristics"

https://link.springer.com/journal/10732

and elsewhere.

when the original algorithm gives better results, why do you make the change.

but we can make the change in the original metaheuristics for the following reasons:

1. the domain of the variables manipulated in the problem treated is not the same in the original algorithm.

2. the change in the parameters of the algorithm in order to obtain better results by taking into account the following two principles: diversification and intensification.

In other words, the values ​​of the parameters must not favor one principle to the detriment of the other.

Firstly, I would be very wary of the type of heuristic algorithm you are using.

Many of the algorithms seem to be novel and unique, but often they are no more than a more fundamental algorithm with parameter names changed to seem more "natural" or "nature-like".

I highly recommend that you read:

Metaheuristics -- the metaphor exposed by Kenneth Sörensen

which is available here on ResearchGate.

https://www.researchgate.net/publication/237009138_Metaheuristics_--_the_metaphor_exposed

There are also some algorithms that seem very good at first glance,

but perform very poorly when tested on more difficult, less biased functions.

There are many, very subtle biases that can affect reported performance.

An important paper you should read in this regard is also on ResearchGate:

The Evolutionary Computation Methods No One Should Use - Jakub Kůdela

https://www.researchgate.net/publication/366902789_The_Evolutionary_Computation_Methods_No_One_Should_Use

Some algorithms do well for a small number of dimensions, but very badly for high dimensions.

For example, the Whale Optimization Algorithm does fairly well on the COCO benchmark suite with a small number of dimensions (D=2), but badly as the number of dimensions increases. See:

https://numbbo.github.io/ppdata-archive/bbob/2023/index.html

In short, it might not be that it is your fault the algorithms do not work well with "non-standard" parameters. It could be that the algorithms are flawed in some way, or you are using it on problems that are more difficult than those the designers used in their research papers.

Good luck with your research!

Leo.

To complete the answer of Leo Lazauskas you may also read

Camacho-Villalón, Christian L., Marco Dorigo, and Thomas Stützle. ‘Exposing the Grey Wolf, Moth-Flame, Whale, Firefly, Bat, and Antlion Algorithms: Six Misleading Optimization Techniques Inspired by Bestial Metaphors’. International Transactions in Operational Research, 2022

Molina Cabrera, Daniel, Javier Poyatos Amador, Eneko Osaba Icedo, Javier Del Ser Lorente, and Francisco Herrera Triguero. ‘NATURE- AND BIO-INSPIRED OPTIMIZATION: THE GOOD, THE BAD, THE UGLY AND THE HOPEFUL’. DYNA Ingeniería e Industria, March 2022. https://doi.org/10.6036/10331.

--------

The point is that many algorithms, like GWO and WOA, are indeed biased, because intrinsic mathematical flaws. In particular they often are translation sensitive. So the efficiency is highly depending on the test functions. For example they can be pretty good when the solution point is (0,0,...,0) or near to the center of the search space, but very bad if not.

Hi,

Maybe this work can helps you. It present and compare several ML and optimization algorithm.

Article Hybrid approaches to optimization and machine learning metho...

Thank you to all

It depends on what modifications you make. Assuming that you are trying to improve the overall fitness performance of the algorithm, the main changes can be made in the exploration and exploitation phases of the algorithm. However, these modifications may result in imbalances during the exploration and exploitation phase. The performance of the algorithm also depends on how efficiently the solution vectors and fitness function are designed. Therefore, the performance of your modified algorithm depends on various factors.

If you are interested, you can explore quantum-inspired metaheuristic algorithms such as QPSO, QDE, etc., and see how they improve over classical metaheuristic algorithms.