2-SPRT
Posted: Wed Jan 28, 2015 1:25 pm
I started to run some tests again and just to do something different I decided to use the 2-SPRT instead of the SPRT. Recall that the 2-SPRT optimizes the worst case behaviour of a sequential test which is when the real elo is precisely in the middle of the upper and lower elo bounds. Also unlike the SPRT the 2-SPRT has an absolute upper bound on its running time.
Here is the script I wrote some time ago to calculate the 2-SPRT parameters.
http://hardy.uhasselt.be/Toga/sprt2.py
To validate the results the script can do simulation and it can even compute the operating characteristic (OC) and average sample number (ASN) of the test exactly!!
Here are some notes on the formulas
http://hardy.uhasselt.be/Toga/support_whitehead.pdf
These are mainly notes to myself, so they are neither very rigorous nor very readable. You can see that the formulas used by the script are based on approximations but these hold to a high degree of accuracy for small elo differences.
Here is the script I wrote some time ago to calculate the 2-SPRT parameters.
http://hardy.uhasselt.be/Toga/sprt2.py
To validate the results the script can do simulation and it can even compute the operating characteristic (OC) and average sample number (ASN) of the test exactly!!
Here are some notes on the formulas
http://hardy.uhasselt.be/Toga/support_whitehead.pdf
These are mainly notes to myself, so they are neither very rigorous nor very readable. You can see that the formulas used by the script are based on approximations but these hold to a high degree of accuracy for small elo differences.