As far as I know AIS is population-based. I don't know CRO. I would even say that most (if not all) meta-heuristics in the broad field of computational intelligence are population based. The difference between them is the way to create new solutions and how to select good solutions. The results is that some algorithms perform better, or are tailored better, for certain problems. Example: Both PSO and ACO can be classified as Swarm Algorithms, however ACO is more suited to combinatorial problems, while PSO is mainly used for real-valued optimization. Else I think the that the use of a specific algorithms depends on the personal choice of the researcher. If you are interested in their difference you can look at the results on benchmark test.

Is it possible to say that AIS can also be used for optimization problems as well as the others (PSO, ACO, GA, etc.) ? I'm baffled for what perform better in a specific areas (e.g. optimization) considering the similarity that between all these algorithms possesses. Seems to me that "nearly similar" operators that each of this algorithm possesses may not be a way to develop the "best", since I'm studying the possibility to prove that 'no-free lunch' theorem is wrong; not that I say AIS is the best or anything. But it's possible to say a specific algorithm may be 'suited-well' for optimization purposes. Any comment?

Trying to prove that no-free-lunch theorems are wrong sounds like a very big risk. (Even without knowing the details of your study :) ). On the other hand it is possible that certain algorithms are better suited for a specific problem. This kind of insight is also more useful if you want to write good papers.

I agree that trying to prove that the no-free-lunch theorems are wrong seems very difficult. However, from pure empiricism we know that if you consider problems with common properties (i.e. not all possible problems as done in the no-free-lunch theorems) some algorithms perform better than others. That is the reason why there exist many benchmark suites. And there also exists already some work comparing several algorithms on a given problem on a theoretical level. How far you can do the abstraction of algorithm (concepts) and problem properties, I don't know. But this would very good papers. :)

Partial agreement seems for myself, however proving the no-free-theorem to a certain extend may be possible? Seems that empirically consideration of the problems and abstraction of the algorithms seems to me existed quite clear relation among them. However, still not a concrete gap found yet. Still, inspiration and looking the agenda from different point of view may provide beneficial discovery. Thank you :)

As long as you start from the same assumptions of the NFL Theorem, I seriously doubt one can prove it wrong, in the same way one can't prove the Second Law of Thermodynamics wrong. The NFL is a standard result in machine learning and optimization in which evidence matches what is mathematically postulated.

Now, if you are interested in specific fitness landscapes such as those with polynomial-size descriptions, then Grover’s quantum search algorithm can be proven optimal for unstructured search.

In other words: once you define the class of landscapes, then it is possible to point out a search heuristic that can be proved better than others; but if you consider all possible fitness landscapes, then NFL Theorem holds.

.

unsure if AIS enters in the "co-evolutionary class" of optimization problems but the following paper indicates that free lunches do indeed exist in that class

Wolpert, D.H., and Macready, W.G. (2005) "Coevolutionary free lunches," IEEE Transactions on Evolutionary Computation, 9(6): 721–735

(available through citeseer)

.

These meta-heuristics differ from each other in details , but with similar frameworks. These approaches have the garantee to obtain not-so-bad results or to obtain results better than blind guess, if you apply them only with the frameworks themselves.

However, some works have shown their tiny differences in accuracy, convergence and speed. For example, standard PSO often has lower accuracy but sometimes faster than GA or SA.

By the way, SA might be classified into the category Stochastic Algorithm, since its Population often consists of only one individual.

I notice that even trajectory-based meta-heuristics (i.e. those that optimize a single solution during the search process, such as Simulated Annealing, Tabu Search, etc.) have been extended by several authors to create population-based approaches.

Moreover, there are a large number of hybrid approaches that combine the advantages of two or more methods, which is why it is, nowadays, to establish a commonly accepted classification of meta-heuristics.

It could be very difficult, if not impossible, to disprove NFL in the general situations. I agree with most of the previous comments that some algorithms are more accurate than the others at a given landscape. Moreover, the parameters of the meta-heuristics can to be fine tuned to fit different situations. For example, the mutation probability in GA controls the diversity of the solutions.

In line with the second comment by Steffen Finck, I would like to mention the work done with the so-called hyper-heuristics.

Using the words of Peter Ross (see first link), "[hyper-heuristics] might be thought of as heuristics to choose heuristics".

http://www.inf.ufpr.br/aurora/disciplinas/topicosia2/livros/search/hyper.pdf

http://www.cs.stir.ac.uk/~goc/papers/ChapterClassHH.pdf

@Nima, in practise, problems that can be solved within 40/100000 seconds are not really interesting. Many of the problems I'm solving take a few minutes for each evaluation. A single run to convergence may take weeks. An algorithm that can solve it within 1 week is defnitly better than one that takes four.

I do agree with Yan-Fu that choosing the right parameters heavily influences the outcomes. Its probably easy to show that for a certain landscape some parameters are better, while for a very similar one, other parameters might be better -> NFL.

@Nima, then we agree

it may be helpful to you

http://www.academia.edu/3589014/A_Review_of_Population-based_Meta-Heuristic_Algorithms

I wish to know more about chemical reaction optimization (CRO) algorithm.

https://www.researchgate.net/publication/296636431_Codes_in_MATLAB_for_Particle_Swarm_Optimization

Code Codes in MATLAB for Particle Swarm Optimization

https://www.researchgate.net/publication/297245624_Particle_Swarm_Optimization_Algorithm_and_its_Codes_in_MATLAB

Research Particle Swarm Optimization: Algorithm and its Codes in MATLAB