# 130

Dear Igor Sbovoda and Dmytro Lande

I read your paper

Enhancing Multi-Criteria Decision Analysis with AI: Integrating Analytic Hierarchy Process and GPT-4 for Automated Decision Support

My comments

1- First of all, l was pleasantly surprised by the subject of your paper; it appears to be the first, at least for me, to incorporate AI in MCDM

2- In page 2 you say, referring to AHP “This method was designed to address complex decision-making scenario”

Not in my opinion. It was designed according to a lineal hierarchical structure that is not adequate to model a complex scenario.

The pivotal moment you refer is inexact. At the time of its appearance there were other MCDM methods addressing complex problems, starting 1n 1948

You say: “providing a rigorous framework for decomposing decision problems into a hierarchy of more easily comprehended sub-problems, each of which can be analyzed independently”

Breaking a problem is useful to understand it, but not for solving, and I am afraid that they cannot be analyzed independently because normally they are related according to the Systems Theory, and thus, they are normally interconnected. It is like for designing a car you study independently the engine, the transmission, the aerodynamics, the electric system, etc., but all of them are related. The engine depends on the aerodynamic of t he body, depends on the electricity consumed by the car and from the mechanics of the transmission, that is the gear box and cardan shaft

A MCDM problem is a system, and as that it cannot be solved adding up the best results of each part. The whole is generally no equal to the sum of the parts.

3- You say “specially designed format that facilitates a forced choice paired comparison, enabling decisionmakers to systematically evaluate the importance of each criterion against others”

Pair-wise comparison is not normative but descriptive, and as that irrational. On what grounds can you assert that criterion C3 is 3 times greater than criterion C1? Only on the DM intuition, mood or imagination. Nothing scientific indeed

4- “Its unique capability to merge mathematical precision”

How can you say that a method based on intuitions and feelings, has mathematical precision?

5- As you said, the GPT is trained at large data, which is the heart of AI. However, even when your idea of using GPT is indeed revolutionary and very interesting, apparently it does not take into account that in MCDM, each problem, even on the same subject of another one, can be different.

And this is because it is true that you can get valuable information on the type and number of criteria using GPT, but don’t forget that in real practice those are determined based on the series of alternatives they will must evaluate, and as you know, this is opposite to the AHP procedure, that first selects the criteria.

The GPT most probably can give you information linking criteria to alternatives, but it is for me unprovable that those may coincide with your case.

Consequently, if a company decides to analyze three different projects, the DM must know first which are the projects, and then, go to the technical person that in the company is responsible for each area, as engineering, finances, costs, etc. It is them who know the problem and can tell the DM which is the criteria that for each individual point of view should be considered. It is from them, after many meetings and discussions, that normally take into account personal interests, that the final set of criteria is established. This is something that the GPT cannot do.

You have to work with a unique organization, not with the average of 1000, because the number of stakeholders, their ideas, their needs, their wishes and their interest are not condensed in a set given by GPT.

Remember that in despite of the tremendous experience we have on agriculture, we are still unable to predict the volume of the next harvesting or even predict when it is going to rain in adequate amounts for crops

6- Something that you did not mention is that AHP demands independent criteria, something that practically does not exist in the real world. How do you query GPT about this? Will you ask it to select criteria that satisfy this condition? If you do not input your alternatives, how it will know that must adjust to such stringent condition? , albeit you did not to worry for it if you work with ANP.

It appears from your questions to GPT that you are maximizing everything. How do you proceed if you have to minimize a criterion like ‘cost’, together with 10 or 15 more criteria?

7- In page 6 you do not detail detail the final criteria. Curiously, don’ you think that communications as well as back up systems are missing? Common sense tells you that that is fundamental

8- In page 6 you asked a machine for subjective values? How can a machine do that? It only got a set of real values and it is unable to intuit anything; that is for human and may be for animals.

9- In gage 13 you talk about matrix consistency. Could you please explain why the estimates of a person must be consistent or transitive? Who says that?

And most important, is there an axiom or theorem that demonstrates that what is the mind of persons is replicable in the real world, which is in general intransitive?

I hope my comments may help you

Nolberto Munier