8

In the paper that describes how ChatGPT could pass an MBA exam there is a question and associated answer relating to queueing that seems intuitively wrong. I copy the question and the answer the author of the paper gives (NOT ChatGPT) in full below, in short the question is: given a polling station with a max rate of 12 people/hour and an expected flow rate of 10 people/hour what is the average length of queue experienced, with an answer of 5 people. With max rate higher than the expected rate I would expect an average closer to zero than five, and I would expect the average rate to depend on how long the station is open, as the first person is sure to have no queue.

I looked at the wiki page on queuing theory, and I could not find anything that looks like "Average Waiting Time = Average Processing Time x Utilization / (1-Utilization)".

Question (written by exam authors)

The Pennsylvania Department of State is implementing a new electronic voting system. Voters will now use a very simple self-service computer kiosk for casting their ballots. If that kiosk is busy, voters will patiently queue up and wait until it is there turn.

It is expected that voters will spend on average 5 minutes at the kiosk. This time will vary across voters with a standard deviation of 5 minutes. Voters are expected to arrive at a demand rate of 10 voters per hour. These arrivals will be randomly spread out over the hour (you can assume that the number of voters arriving in any time period follows a Poisson distribution).

What is the average amount of time that a voter will have to wait before casting their vote?

Answer (written by paper authors)

To find the right answer, one must look at a standard equation from queuing theory. The equation for the average waiting time states that:

Average Waiting Time = Average Processing Time x Utilization / (1-Utilization).

Plugging in an average processing time of 5 minutes and an average utilization of 5/6, we get:

Average Waiting Time = 5 x (5/6) / (1 - 5/6) = 25 minutes.

So, the correct answer is 25 minutes waiting in line. If we add the 5 minutes at the kiosk, we obtain a total of 30 minutes.

User65535
  • 197
  • ChatGPT became famous for no reason. It cannot do math and is not very useful for anything else. – Peter Mar 07 '23 at 09:36
  • 9
    There are currently two close votes on this question, both because the question is "not about mathematics". I feel like these voters did not read the question carefully: the question is not about ChatGPT, it is about queueing theory, in particular about an MBA question on queueing theory that OP found in a paper about ChatGPT. Queueing theory is unambiguously on-topic here. – Mees de Vries Mar 07 '23 at 09:57
  • 2
    I am one of the close voters. Willing to retract the close vote as soon as any reference to chatGPT got removed. Here in MSE we have ample evidence that chatGPT has nothing to do with mathematics. @MeesdeVries – Kurt G. Mar 07 '23 at 10:27
  • 10
    @KurtG., this question is no more about ChatGPT than the moving sofa problem is about furniture. The authors of a paper about ChatGPT claim that the second quotation in the question is a correct answer to the exercise in the first quotation. The question is whether or not that is true, and that is a question about queueing theory, not about ChatGPT. OP just gave the context for their question (which they are required to do!). – Mees de Vries Mar 07 '23 at 10:36
  • 1
    @MeesdeVries I copied and pasted that question from the paper OP linked into my chatGPT account and after a very long saga it came back with "So on average, a voter will have to wait approximately $\color{red}{4.4}$ minutes before casting their vote." This is evidence that chatGPT produces inconsistent results every time it is being asked the same question. If the question is solely about queuing theory it should be edited. – Kurt G. Mar 07 '23 at 10:52
  • I have edited the title to remove ChatGPT, and highlighted that nothing in the question is from ChatGPT, the quotes are from the authors of the exam and the paper. – User65535 Mar 07 '23 at 11:03
  • 1
    This question is clearly about ChatGPT , namely whether it would give this answer and therefore pass the exam. So, in the current form , it is off-topic. – Peter Mar 07 '23 at 11:18
  • 8
    @Peter The answer quoted is given by the paper authors to allow people to assess the answers that ChatGPT gave. The only relevance ChatGPT has to the question is the the paper where I read it was about ChatGPT, no ChatGPT generated content is in my question. – User65535 Mar 07 '23 at 11:27
  • 1
    Posted an answer. We do not have evidence that the paper that is still quoted in OP shows that chatGPT did in fact get the right answer because those answers cannot be reproduced. The ability to reproduce is the main paradigm of every exact science. – Kurt G. Mar 07 '23 at 13:23
  • 6
    @KurtG. it seems you are still under the impression that the second quotation in the question is ChatGPT output; it is not. None of the text in the question is ChatGPT-generated. This question is not about the output of ChatGPT. The first quotation is from an MBA exam. The second quotation is written by the authors of a paper about ChatGPT. – Mees de Vries Mar 07 '23 at 14:08
  • 2
    @MeesdeVries . Thanks for clarifying that. This seems the case. I am so put off by that hype that I ditch every text that mentions chatGPT more than once immediately. Learned some queueing theory today and found a terrific simulation done by some hard working chaps. – Kurt G. Mar 07 '23 at 15:50
  • 1
    A typo...? I suppose the last words of the first para in question should read 'until it is their turn', not 'there turn'. Please let me know if I'm wrong. – CiaPan Mar 17 '23 at 10:25
  • Intuition may say that if the service rate is above the arrival rate then there will little or no queuing. Sadly, Intuition fails as the system approaches capacity. Anyone who has managed a large multi-user computer system will know this. – badjohn Mar 17 '23 at 11:05

1 Answers1

3

Not being a queueing theorist I give it a go:

  • The system is an $M/M/1$ queue with mean interarrival time of $1/\lambda$ being six minutes ($10$ voters per hour) and mean exponential service time of $1/\mu$ being five minutes.

  • With the notation $\rho=\lambda/\mu\approx 0.83333$ and by Little's law the expected waiting time of a voter in the system (queue plus casting the vote) is then $$\tag{1} \mathbb E[S]=\frac{1/\mu}{1-\rho}=\frac{1}{\mu-\lambda}=30\text{ minutes }. $$ The mean waiting time (in the queue) is $$ \mathbb E[W]=\mathbb E[S]-1/\mu=25 \text{ minutes }. $$ Formula (1) has an intuitive explanation here.

Kurt G.
  • 14,198