Dear Michael

I am not very knowledgeable in metaheuristics, and therefore, my opinion does not carry any weight.

However, in reading your remarks I feel that I agree with you. Perhaps some people use an AK47 just to kill a mouse, because they do not know about and old a simple device: the mouse trap.

I am a 100% with you on the matter that there are thousands of real-life scenarios that can be solved by MCDM heuristics, although they don’t allow modelling reality as it should, and thus, using a high degree of flexibility and ignoring many actual and important facts, pertaining and being closely related to scenarious.

Many of these problems work with a mix of quantitative and qualitative criteria as well as a mix of linear equations and integer programming, however, using the Simplex there are a few software such as Solver (in Excel) or Lingo, that can find optimal solutions, even with this mix.

In my opinion, some practitioners use actual MCDM methods because that are easy to understand and because apparently very few people know about mathematical programming, which curiously was the origin of MCDM method. In other words, they apply the law of least effort!

I also agree with your third paragraph, because there are very complicated problems that as you say, they do not have an explicit formulation. I would say that AI problems fall in this category.

In your fourth paragraph you say: Yet there are plenty of scholars - especially in this forum, for a reason I do not fully understand - that insist on applying their favourite metaheuristic(s) on just anything.

This is exactly to what I said many times, asking why some colleagues are using a method such as AHP which is deeply mathematical and procedurally flawed. I even challenged people several times to rebut what I said. Even when my last challenge accumulates more than 180 reads, NOBODY answered.

Why?

Because nobody can demonstrate that what I said related to this method is a non-sense, not even its creator. The problem is that it is so easy to reason and show that a method has serious drawbacks, that I wonder why people don’t see them. For me the explanation is that they prefer to ignore what common sense show.

You and I are together in the same fight although in different fronts. You, with the undue use of the meta-heuristics, and I with the ignorance and blatant unawareness of the real world.

Thanks for the very helpful support, Nolberto! We need to stand firm against pseudo-science, wherever we see it. I am not altogether against metaheuristics as such, but wish that they would be used with care, and on the right sort of problems for that particular set of techniques - instead we see hundreds of variations of rather few easy techniques, in which NO method variation can guarantee much at all. And I find it even more puzzling that they tend to shy away completely from mathematical optimization (MO) - as if MO were some kind of plague! I would not yet compare it with the worst cases of pseudoscience, like homeotherapy, but the users of metaheuristics need to be more aware what to expect. We would not want that technique to infiltrate serious sciences like medicine, would we?

Dear Michael Patriksson,

I have previously agreed with your point of view and expressed my opinion at https://www.researchgate.net/post/Is_the_area_of_metaheuristics_moving_away_from_scientific_rigor . However, it is necessary to correctly understand the current situation with science, which has long been transformed into a business. From this point of view, Metaheuristics has long been a part of this business. I personally saw this when I took part in the International Conference MIC 2011 in Italy, Udine, July 2011. I left this conference with great disappointment, because my expectations in order to discuss were broken. There were no discussions at all. Now I have a specific question to you. If you have the opportunity, please go to the International Conference MIC 2019 https://mic2019.uniandes.edu.co/index.php , which will be held July 28-31 in CARTAGENA, COLOMBIA and share your opinion. Especially it would be very valuable for you to meet the father of metaheuristics, Fred Glover, and express your opinion. However, I am afraid that things have gone so far that these conferences are business, business and again business and it will be very difficult to prove your point.

Dear Gennady

Most of these conferences are like the old society standards; people go there to see and be seen.

I admit that they are very useful to make connections, but other than that, I don't think that too many people are interested in technical discussions and to learn something. Honestly, I believe that Research Gate is by far more useful, instructive, and free.

You said that you were disappointed, and I can understand you, because I have maybe attended and participated with papers, in no less than 15 international conferences, and I left them with the same feeling you had.

In some of them I even proposed to incorporate a workshop to discuss subjects of interests. Needless to say, I was not heard. That is not business@

I have a very different experience from my math conferences; debates are all over the place - not just about the strength of the research, but immediate comments on errors, or "what you wrote can be extended very far - why haven't you?" or "I can prove this much more elegantly". But that is rather rare these days, yes.

Michael

Perhaps I didn't explain myself very well.

I believe that we have to make a difference, that I didn't, between international conferences, with their usually heavy fees, and those seminars, workshops, and conferences that take place when a scientist is invited by a university.

The first type, has two parts. The first part, for me at least, are very agreeable social gatherings, with good food, abundant coffee and pastries, excursions, and a lot of speeches by prominent people saying how good is to have so many people from many difference countries, and that this gathering was a step to a future and a long and useless speech, that lead to nothing.

As you know they are called the 'Key speakers'.

A second part is a ten or fifteen minutes, or if you are lucky, 20 minutes long dissertation, about the use of some MCDM method in a project, and usually, pointing out how successful it was.

There are five or perhaps 10 minutes for questions and answers, and that is, and you are lucky if you get a sizable number of people, which also depends on the day and hour of your presentation, because people get bored.

In my last presentation in Canada, two years ago, I was one of the last presenters in the evening of the third day of the congress. I had six people attending my speech, and only one asked a question, and I think that it was out of pity....

The second type, is when a scientist is invited by a university not to give a speech, but to address a specific subject, and most important, when the gathering involves a workshop, with a duration of hours or even days, where most attendees know about the subject, because the dissertation was sent to them before the event. In this way they are prepared to discuss with the presenter, argue with him/her and perhaps reaching conclusions.

I don't know to which type the math conferences you gave belong to. However, if an international conference about several aspects of the same subject, lasts only minutes, I wonder which is the reward to your effort and which the benefit to the attendees.

People asking questions after the exposition are pressed to hurry up because the room is needed for another conference. Sometimes I tried to talk with the presenters after the dissertation, and I was blessed with four or five minutes talk. In this case the presenter may be or not in agreement with you, but what is the balance?

Naturally, I respect your opinion, and it could very well be that I am mistaken.

I think that this is a good subject to be debated in in Research Gate, asking for people's opinion about the so called scientific international conferences.

Nolberto Munier: The very best conferences are small and are placed at an isolated address, such that the only ones you can talk are the attendees - and they should not exceed 30 or so, counting also the conference organizers. The ones in Erice, Taormina, and a few others, on Sicily, located way off the major roads, and I am sure also in Greece, are wonderful and relaxing, and you have time to talk with the attendees, as there is not much else to do, having done the ususal - lovely, yes - excursions, and dinners. That's my kind of workshop! :-)

Dear Michael Patriksson,

You raised a very important question, but I got the impression that something is stopping you from developing this question, perhaps some external factors like “Big Brother”, which controls us everywhere and everywhere. Why do I have such an opinion? Because despite all the attempts to call metaheuristics as a pseudo-science, we see in reality many magazines and conferences that generously spread the ideas of metaheuristics. Do they not understand the pseudoscience of this subject? I think that they are well aware, but do not want to notice.

Let's see https://mic2019.uniandes.edu.co/index.php/committees . I am personally acquainted with the members of STEERING COMMITTEE:

FRED GLOVER

University of Colorado and OptTek Systems (USA)

CELSO RIBEIRO

Universidade Federal Fluminense (Brazil)

ÉRIC TAILLARD

University of Applied Sciences of Western Switzerland (Switzerland)

STEFAN VOSS

University of Hamburg (Germany)

Do you think it is worth or not to write a letter to these people about your vision of this problem. If these people deliberately spread false ideas, then they should know about it.  What is your opinion?

Gennady: There are a few problems, it seems to me:

The workers in the field of metaheuristics do not mingle much with the mathematical optimization scientists, and they certainly have a problem answering questions. One may ask why, and the answer may be evident, in fact. But since you have a good connection with them, could you ask a few of them why they isolate themselves, if that is what they do.

We would very much like to see more interactions, if possible, perhaps make more comparisons among them such that we can learn when metaheuristics may make sense, or find better ways to combine them, if that is fruitful for some problems.

i'm sure that in for example infrastructure engineering and similar "practical" OR there is plenty of interaction among engineers with lots of knowledge of math that collaborate. Perhaps such projects are already running, but I do not have the connections such that I could know the status. There certainly should be possible for metaheuristics to play a role, if not to find an optimum, but to explore parts of space such that better and more focussed approaches can "take off" at regular times. Or something like that - the options are endless.

Dear Michael Patriksson,

Not a day later, a new discussion appeared https://www.researchgate.net/post/What_is_the_scope_of_earthworm_optimization_algorithm_EWA_in_engineering_optimization_problems_compared_to_other_existing_optimization_techniques .

Long live the "earthworm"!!! No, metaheuristics are eternal. Well, since people like it, then nothing can be done, let it be so.

P.S. Article Earthworm optimization algorithm: a bio-inspired metaheurist...

I have a few new names that RG meta-heuristics people may borrow:

1. The bored panda algorithm,

2. The amputee algorithm, and its more modern implementation,

3. The buried corpse method, followed by

4. The rotten apple devise (RAD), and, the best one:

5. The SCAM (Sociopath Catastrophe Ambient Method).

I hope they will all be very worm-while - sorry, worth-while -, and lead to many new publications.

Michael, metaheuristics are just practical solutions to practical problems. They work just fine as long as you don't get picky with precision. If you wanna go mathematical about it, you'll quickly find that proving stuff like convergence of the algorithms to be quite the complication, which is why such proofs are very rare.

As to why people suggest them as methods to solve simple problems... well, that's normal: people tend to recommend what they are familiar with.

But precision is key! I don't want a quick-and-dirty solution. I want a proven optimum, whenever I can. And I use tools that (most of the time) actually give me pretty good estimates of the optimal value of the problem from both below and above.

Why would one ever generate a "result" to a problem, when you haven't got a clue how good it is? It's absurd. It's definitely not practical (and that was, I think, the original purpose of devising these heuristics), especially if you are dealing with a serious problem in industry, and not being near-optimal means spending huge additional resources as a result. You could go bankrupt.

Well, I am sure you can find many real-world examples where near-optimal results do just fine. The industrial and corporate world usually abides by the Pareto law "spend 20% of the resources to get 80% of the solution" because time is a factor as well. Spending more and more time to derive precise solutions delays time-to-market and that is a big no-no. You only do that if you really need to.

Still, I do agree that from an academic point of view it is important to prove optimality and convergence of metaheuristics. I've never been into that line of research, but it's a very interesting one. Even if I feel most results are likely to be negative. I.e. "this metaheuristic does not converge in good time".

Dear Michael and Nuno

Very interesting and productive discussion between you two and in my opinion, both are right.

I like Michael sentence ' Why would one ever generate a "result" to a problem, when you haven't got a clue how good it is?

It is in line with what I have been expressing here in RG, and I can't understand how a practitioner 'solves' a problem using any MCDM method, and then happily says that he was successful. In reality he/she may be indeed be successful in feeding an algorithm with data and running it; it proved that the method can deliver a solution considering the data inputted. Not a big deal!

But how a person can say that he reached a good result if he does not know which the 'real' or 'best' result is, or even is a solution exists?

Obviously, there is not a yardstick for results, and if it were one, then, no one would need MCDM methods...

Now, if he uses a method that can guarantee optimality, then, he can claim that it is believed to have reached the right solution, because mathematics don't lie.

Your sentence Nuno: 'Well, I am sure you can find many real-world examples where near-optimal results do just fine'., is also true, and it is the foundation of all heuristic methods. Problem is, that you never know when there is a near optimal solution.

All heuristic methods assume that because reasoning, and using well math founded algorithms, they can reach a near solution or the solution. In reality, what they reach is an average of opinions and wishes, or a consensus among acting DMs.

The Pareto 'law' of 20/80, is a very valuable tool, but that unfortunately does not have any mathematical support. It is pure heuristics, which has the virtue that most of the times it can be checked with reality, and nobody knows why, usually it is correct.

It is also true that in many occasions you can't afford spending a lot of money and efforts looking for better or optimal solutions.

Because of that I believe, since at present at least, we are unable to find an optimal solution, if it exists, that heuristic methods are a good approximation. We don't have other choice!

The problem is, that working with optimal or heuristic methods, if the scenario is not properly modeled, all efforts are in vain

I also like this Michael sentence 'There certainly should be possible for meta heuristics to play a role, if not to find an optimum, but to explore parts of space such that better and more focused approaches can "take off" at regular times', and believe that they may establish a bridge between optimality and meta heuristics.

A friend just told me not to answer (even if it is only now and then). He supposed that even good answers do not change people's mind. Anyway, we had these discussions again and again. But these discussions seem to get lost every now and then. So, let me bring it to one or two threads again, may be using different words.

Like in every field there may be good research and bad research. So, do not put all of it together into one pot with possibly even wrong comprehensions.

Regarding the metaheuristics community, there is some awareness at least in parts of it that not everything that seems shiny is shining. A good reference to explore this view is by our friend Kenneth Sörensen:

Metaheuristics—the metaphor exposed

https://doi.org/10.1111/itor.12001

And another good pointer is to the word matheuristics which we framed to investigate the interoperation of metaheuristics and mathematical programming techniques. Is it allowed to point to a wikipedia page? May be yes: https://en.wikipedia.org/wiki/Matheuristics

Have fun exploring and learning how to separate the wheat from the chaff :-)

Dear Stefan Voss,

I listened to your report on July 27, 2011 in Udine and now thank you for joining this discussion on the threshold of MIC 2019, where you are an important person from the conference management.

Currently, I am not a big fan of metaheuristics and do not consider it a science at all, keep this in mind. My main complaint is the absence of any attempts to find the Lower Bound with which to compare the heuristic solution. The search for such Lower Bounds is a real science, sorry for being direct. However, this is not done. Because the search for good Lower Bounds requires fundamental knowledge, but this fundamental knowledge does not need at all in order to find any heuristic solutions. Soon there will not be a single insect and mammal, which would not be used to simulate its behavior for the publication of the next metaheuristic algorithm, of which there are already hundreds of thousands, if not more. However, the search for good Bottom Bounds is a vacuum. I set forth my views on Metaheuristics at https://www.researchgate.net/post/Is_the_area_of_metaheuristics_moving_away_from_scientific_rigor

However, now I would like to offer a different perspective on the problem https://www.researchgate.net/post/Should_be_revised_todays_Complexity_Theory

I would be grateful if you refute this view, but please do it with any arguments.

Thank you in advance.

In addition to the starting points posted by Stefan Voss, I'd like to mention that some journals are enforcing policies for more rigorous research practices in the field.

For example, the Swarm Intelligence journal (see the "Additional submission instructions" at https://www.springer.com/computer/ai/journal/11721) and the Journal of Heuristics (PDF: https://bit.ly/2Q4gBCx) have such a policy.

Bio-inspired metaheuristics are fine. The tendency to design metaheuristics reproducing every biological system is however probably overcooking it. But to me this is more a consequence of the "publish or perish" lore than anything else. I don't blame researchers who need to build a career and go this way.

The problem with the papers - or for that matter, the whole journals, that are devoted to metaheuristics - is that they have criteria that are much less demanding than the classical ones; in other words, I'm saying that the published papers in the area of metaheuristics are NOT scrutinized based on the same, tough, criteria by which classical optimization/numerical analysis papers are. For example, mathematical optimization papers dealing with numerics typically demand quite thorough numerical comparisons on (often [semi-]real) large-scale cases, coupled with performance measures that are defined based on actual optima. Metaheuristics have none of that, only small examples pulled out of a very dirty hat, like a drunk magician at a worn-down circus show in the suburbs.

Just because these "scientists" need to publish does not mean that we need to dumb-down the journals to cater their needs, do we?

That's why journals are divided into quartiles. Only the good stuff reaches top journals. Supposedly... :)

;-)

Dear Michael Patriksson,

"But since you have a good connection with them, could you ask a few of them why they isolate themselves, if that is what they do".

Here in this discussion I asked one of them as Stefan Voss and so far there is no answer. However, in anticipation of an answer (if of course this answer comes sometime) I want to once again designate my position in a popular form. So, you want to find the optimal (exact) solution OPT(D) of some combinatorial optimization problem, where D is the initial data. Consider, without prejudice to the generality, the minimization problem of the problem. Then any approximate (heuristic) solution A(D) >= OPT(D). The problem is finding the quality A(D) compared to the optimal solution OPT(D), which is unknown. The value q = 100% * (A(D) - OPT(D)) / OPT(D) is taken as the measusre of quality. Since OPT (D) is unknown, the Lower Bound LB(D) is considered, where LB(D)

Dear Gennady,

Instead of demanding an expert answer from Stefan (who has probably many other important things to do besides these forums), I recommend you to consult an introductory course on computational complexity. Most problems of interest have inapproximability results. For the set covering problem for example (frequent as a subproblem in many industrial applications), no deterministic algorithm can return a solution within a factor o(log n) of the optimum in polynomial time unless P = NP. In such a situation, you have to choose between 1) solution quality guarantees or 2) tractability. In most application cases, practitioners choose option 2). Moreover, I suggest to consult some recent papers on (meta-)heuristics from either Journal of Heuristics, INFORMS JOC, Transportation Science, or Operations Research, among others. In their wide majority, these articles contain a section reporting lower bounds or exact results to allow comparisons with heuristic results on some small (solvable) instances. As these instances are typically too small to make meaningful conclusions and analyses, additional comparisons are made on larger instances with previous methods as well as with the entire history of solutions (BKS) known for the considered problems. As a matter of fact, researchers in combinatorial optimization and operations research often combine an expertise in exact methods (branch-and-bound, branch-price-and-cut) and in metaheuristics, as it is increasingly common to use metaheuristics components in exact methods (e.g. for cuts separation, pricing, choosing branches or generating primal solutions), and exact components in metaheuristics (e.g. for solving subproblems).

Dear Thibaut Vidal,

"In their wide majority, these articles contain a section reporting lower bounds or exact results to allow comparisons with heuristic results on some small (solvable) instances. As these instances are typically too small to make meaningful conclusions and analyses, additional comparisons are made on larger instances with previous methods as well as with the entire history of solutions (BKS) known for the considered problems".

We talk about different things and it seems to me that our conversation is a conversation between the deaf and the blind. I mean one thing, and you mean a completely different thing. This is not about the fact that a heuristic solution can produce very good solutions, but it’s about the fact that the heuristic solution does not know anything about it, I want to especially stress, it does not know anything. I have repeatedly mentioned that the search for a heuristic solution in itself does not mean anything until the quality of this solution is clear by the time the search for a heuristic solution is completed. In other words, as soon as you receive a heuristic solution, you should immediately declare the quality of this solution. Not later, but now, right at the moment after the end of the algorithm. Lower Bounds need to be found not for small examples, but for large examples for real problems. Now I propose to play the next hypothetical game. Suppose someone X announced a prize for solving a combinatorial problem within a given time limit, for example, one hour. There are two algorithms: You represent one heuristic algorithm, and me represent the other heuristic algorithm. X claimed: the prize will be received by the one of you who is the first to declare that its heuristic decision has the quality of a solution q >= 80% (about "q" see above). Your algorithm will produce solution A1(D), and my algorithm will produce solution A2(D). My algorithm has the following strategy: Half the time limit, that is, 30 minutes, I will spend to find the heuristic solution A2 (D), and the next 30 minutes I will spend to find the Lower Bound LB(D). In other words, your algorithm runs for 60 minutes, and mine runs for 30 minutes, but I’m finishing work with the LB(D) found. Now suppose that your algorithm got a much better solution than my algorithm, that is, A1(D) < A2 (D). That is, the quality of the decision you won a prize. However, if it turns out that p = 100% * (A2 (D) - LB (D)) / LB (D) >= 80%, this means that by the end of the contest I will declare that I was the first that my algorithm received a solution that satisfies the requirements of X. That is, despite the fact that your solution is better than mine, I get the prize, not you. This is exactly what I propose to use the strategy: within the specified time limit to declare about the measure of proximity p%.

You probably know the FFD algorithm for searching for a heuristic solution in the One-Dimensional Bin Packing Prroblem. The classical D.S.Johnson's theoretical theoretic result FFD (D)

Dear Gennady,

The cornerstone and the flaw of your reasoning is the following: "Not later, but now, right at the moment after the end of the algorithm. Lower Bounds need to be found not for small examples, but for large examples for real problems". It is surely an admirable requirement, but it is impracticable for most problems of interest, at least in the current state of knowledge. Consider a minimization problem. There is a huge asymmetry between finding a primal bound (i.e., a feasible solution/UB) and proving that no solution better or equal to a value X exists (i.e., a dual bound/LB). Metaheuristics are remarkably effective for the former (UB) problem, but the latter problem (LB) typically requires extensive (exponential-time) branching and bounding... Worse than that, you often face pathological situations when generating LBs: dual bound getting stuck close to the trivial value (i.e. zero) basically forever, symmetry causing excessive branching, or inability to even complete column generation at the root node (for branch-and-price methods). In the overwhelming majority of cases, on large-scale problem with a difficult structure (I'm obviously not talking of the knapsack here), the LB you can obtain within acceptable CPU time is too far from the optimum (e.g., 50% or more) to give any meaningful information or performance guarantee. And this is not from a lack of research: nearly 70 years have passed since the seminal works of Dantzig and dozens of thousands of studies have been conducted on the subject.

If you are not convinced by these comments, I recommend, as a useful exercise, to attempt to model and produce good LBs for a Quadratic Assignment Problem, or an Inventory Routing Problem with 100 nodes.

Dear Thibaut Vidal,

I am glad that you are interested in the LB search problem and now I have a good reason to develop a discussion on this issue. First, why do you think that only LP-programming can be used to search for LB, although this sometimes gives a good result. However, for problems with large sizes, this method is not applicable in practice. In my understanding, it is necessary to develop special methods that would make good use of the nature of the problem and the source data. I developed such special methods for my problem. Now about the Time Limit. First I try to find the Initial Trivial Lower Bound LB(D, 0). Then I use the "branch & bound" method to solve the problem of whether "is it true that OPT(D) > LB(D, i)?", where LB(D, i) is a predetermined threshold, LB(D, i) > LB (D, i - 1) > LB(D, i - 2)> ...> LB(D, 0)" within the specified time limit. Here, LB(D, i -1), LB(D, i - 2), ... LB(D, 1), LB(D, 0) are the previous lower bounds found by the time when the lower bound LB(D, i) is searched. If it is not possible to answer the question “is it true that OPT(D) > LB(D, i)" within a specified time limit”, then this threshold is lowered, for example, to LB (D, i) - 1 > LB(D, i - 1). For many models, it was possible to find a series of first LB(D, i), i >= 0 in a very short time in order to compare the found heuristic solution A(D) with the last found value of LB(D, i). If the condition p = 100% * (A (D) - LB (D, i)) / LB (D, i) >= (given threshold) is satisfied, then the process is interrupted, because the measure of quality of heuristic solution is satisfied. That is, it is not need to achieve the result LB(D, i) = OPT(D). However, the last found Lower Bound LB(D, i) could become the final LB(D) = LB(D, i) within a given Limit Time, for example 30 minutes. Other special methods of searching for qualitative Lower Boundaries are possible, which would well take into account the nature of the combinatorial problem and the nature of the initial data. However, finding these methods is hard work.

Proposing a “new” metaheuristic is very tempting. By new I mean here its actual meaning. In my opinion, one of the reasons that has not been mentioned here by metaheuristcs supporters is beauty. Indeed, I feel a great pleasure to contemplate the ideas behind some clever metaheuristic, especially for genetic algorithms. Unfortunately, nowadays pseudo-new metaheuristics are simply a medium to quickly boost one's citation index and to increase in the way journal's impact factor.

Your final sentence is so true, Menouar. Spot on!

I've already seen a paper on genetic algorithms which were "almost" indistinguable from a paper on biology. It was seventeen years ago!!!

It is true there existe nowdays several proposals which can be considered almost copies from other methods

New discussion

https://www.researchgate.net/post/How_can_you_differentiate_Science_from_Pseudoscience_in_your_social_and_academic_life

could be a good complement to our discussion about Metaheurictics as Pseudoscience.

Thin ice!

Pseudoscience is "A pretended or spurious science; a collection of related beliefs about the world mistakenly regarded as being based on a scientific method or as having the status that scientific truths now have." Oxford English Dictionary.

Be careful not to mix the necessary distinction between good science and bad science (that we need to care about also regarding metaheuristics) with pseudoscience.

I just also added an answer to: https://www.researchgate.net/post/Metaheuristics_for_the_simplest_combinatorial_problems-why

Please, do also think out of the box before discussing.

If you want to include me in the discussion, please be more open-minded.

Again, I do not want to defend anybody or anything. I just want to sharpen your eye on different things. It is not always only black and white.