Hello,

I am doing an academic project in which I aim to program intelligent agents, whose reasoning more or less copies human reasoning. The example I am working on is a very simplified stock-market type negotiation procedure.

My problem is that it is not easy to say how human beings make decisions in such a setting, so even harder to create the artificial agents. Therefore, what I would like to ask, is if there would be people willing to help me by filling an online survey, with a couple of scenarios in which you will be asked to say how you would react as a human being, and analyse why you would do so.

I do not need hundreds of answers, but I do need really detailed explanations about your choices to be able to use them. So I would rather have you answer just 1 or 2 scenarios with lots of details than just tick the boxes to give all your answers, but without any justification.

If you have any comments, any suggestions, any questions about the survey or my project, I would be pleased to hear them.

Thank you in advance!

Yes sure, game theory.

But no one has studied my example yet, and there is no all-covering game theory theory that I could instantiate for my own problem. So this online questionnaire is actually there to help me build a theory for my example.

I just get a blank page when I go to your page...

I have a problem with your survey, as you dictate what I am feeling.
"Other case: I am later in the session 1 / 3, I have been trying to sell A to other agents, who all told me they were not interested. Looking at the display, I know that no object A has been traded yet. I start to worry thinking that I should hurry up getting rid of my A, because the offer seems greater than the need. So I change the lowest price I accept to get for A after [1 - 3] “not interested” answers from [0-20] down to [0-20] ."
The need is not less than the offer! There is one person who is willing to buy A and you have only spoken to a possible maximum of three people, I would not move on the price the day is long yet I am in the first session!! What about the option of buying a B or even another A which you maybe able to get cheap and sell later at a profit.

For these reasons and more I did not complete your survey as I would only partially complete it if I did.

I still seem to be blocked?

Can you try to download the .odt version instead please?

Well, if you know there is only 1 agent interested in A and there is you plus possibly other agents willing to get rid of an A, then the offer might indeed be bigger.

Yeah, I have thought about it as well, but I don't want to get too many options, at least in my first model. Perhaps later.

But even if you don't fill the questionnaire, every comment is welcome! I might still change quite a few things before the end of my project.

What do you want to achieve? Is it a project in psychology/sociology (so possibilities to model humans matter) or an AI one (forget humans, get good results)?

In any case, you do not even tell us anything about distribution of goods (A/B). Usually there is some source of information about it. Surely, it is more like statistical, and such data is interpreted differently by different players (and differently by humans and rational players), but anyway.

I would recommend you to start something better defined, and with less opportunities. Especially if you are interested in human behavior. Have you studied prisoners' dilemma optimization, for example? There is some information on AI championships and their past results, but psychological data maybe less accessible (and more interesting or less interesting - depending on your luck..) Or you want to study concrete examples of your trading problem, like 3+2 situation (when everyone knows it is 3+2). Also, think hard about what you want to get in 1+1 situation. Note that your exact model has forbidden packet deal, and it is not really good in my opinion..

Sorry, I haven't enough to spare...

no, couldn't get that either, maybe it's something to do with tor, it seems to be acting funny lately?

This might help develop what raskin is thinking, but probably not?

rubadub: If you want to study humans, maybe you should set up a small online multiplayer game and view logs. If you find correlations with personality types for those willing to take the quiz, it would also be fun..

Asking humans without giving enough information has some drawbacks -people may believe what they say about their behavior without information, but having any means of getting the information, they will act differently..

too true, and this isn't my game. But a lot of peeps I know will go to all lengths to cover their true intentions and mislead the data gatherer. However as programmers, at most stages we have to consider human behaviour even if it's only to prevent the user from crashing the system, intentionally or not. And also I tried to steer Khaan towards such incidences as the prisoners dilemma, but as he states there is no multiplayer incidence, even though a book I can see to the right of me has an example covering the subject of network traffic routing on multiple domains.

But if you you know all the possible moves, the gains/losses from each move and then what type of personality (or person in a certain state) would probably do, then the demographic makeup of the user set, then you can pretty reasonably determine levels of effect.

But hey it's Sunday morning and i'm just a rambling madman and thats me in a pidgin hole... lol

I must apologise I missed the first sentence. The other problems I had have been outlined in that there is not enough information for example traders will tell you some information about A and B's but will they tell you information about the other objects add to this you do not even mention the other objects in the questions.

If you are going to do a survey then all the information is required:
How many people have told me they have A's,B's etc
How many people have told me they want A's B's C's etc
How many people have I spoken to?
What is the full history of my actions?
...

As has been mentioned a quick mock up for learning purposes is what is really required, get five people together (does not have to be online, could be one machine) and have them play out the sessions recording the information. With the gathered information create an AI bot and then play it out again with four/five people seeing how good it is(whilst still recording the info). The problem which the survey route is that is only covers a small proportion of the many possible scenarios and I fail to see how this little information would be usable as the scenarios may never happen in a simulation.

Thanks to all for the comments, even if I am not always responding to all of them, I am taking note.