Maximum ELO gain per test game played?
Posted: Mon Apr 20, 2015 4:52 am
I'm curious...
With limited computational resources, what confidence bounds should be used to maximize the ELO gain per test game played. I guess to be technically accurate we should maximize the ELO gain per unit of computational resources, but to simplify it we will assume that all computers used are identical and all games require the same amount of time to be played.
I'm curious if anyone has seriously looked at this. I know that during the cold war the US sponsored a lot of math research into this type of problem. e.g. It's very expensive to set up, maintain, and secure an ICBM test range. It's even more expensive to launch 20 or so ICBM's at specified targets. Reducing the number of missiles required would obviously save a lot of time and money. The idea was to minimize the number fired while still being able to determine if the design criterion had been met and to periodically re-certify the fleet as it ages.
Anyway, all I really want to know is what confidence interval maximizes ELO gain per unit of computation. I was thinking that a 95% CI was way too high for those of us that have limited resources available.
Regards,
Zen
With limited computational resources, what confidence bounds should be used to maximize the ELO gain per test game played. I guess to be technically accurate we should maximize the ELO gain per unit of computational resources, but to simplify it we will assume that all computers used are identical and all games require the same amount of time to be played.
I'm curious if anyone has seriously looked at this. I know that during the cold war the US sponsored a lot of math research into this type of problem. e.g. It's very expensive to set up, maintain, and secure an ICBM test range. It's even more expensive to launch 20 or so ICBM's at specified targets. Reducing the number of missiles required would obviously save a lot of time and money. The idea was to minimize the number fired while still being able to determine if the design criterion had been met and to periodically re-certify the fleet as it ages.
Anyway, all I really want to know is what confidence interval maximizes ELO gain per unit of computation. I was thinking that a 95% CI was way too high for those of us that have limited resources available.
Regards,
Zen