Uri Blass wrote:I think that we can simply repeat the test many times when we finish a test only after we get error that is smaller than a specific number(let say 0.05 minutes)
If we get in 95% of the cases that B needs 1.95-2.05 minutes against A with a possible error of 0.05 minute then our error function seems to be a good estimate for the real error.
Your mathematical skills are probably a lot higher than mine, so please excuse my silly questions below
How exactly would you like to decide about termination? How do you calculate that error, given a sequence of N games with their individual results W(i) and their individual time assignments T(i)?
a) Do you want to take the average of all T(i), and see how many of them are close enough to it, or
b) do you want to take the most recent time value T(N) as a reference, instead of the average?
I assume b) but I am not sure about your intention.
In your first model you proposed not to change the time factor in case of a draw, which poses the question when to terminate if there are a lot of draws in a row. (Maybe this would be solved by applying the error function, but the problem described next would remain open.)
In my model I have focussed on letting the total sum of all game results converge towards 0.5 * N, as opposed to your strategy of letting the time converge. A possible problem of letting the time converge could be that you have no clear statement about the question whether the initial requirement "equal strength" has been fulfilled, and I could imagine a case where you stop based on the error function for the time values but the sum of game results differs a lot from 0.5 * N. One could also define an error function based on my convergence model, so what do you think about it?
Sven