View Single Post
Old 07-03-2013, 03:04 PM   #38
beees
 
Join Date: Jun 2013
Gametype: Capture the Flag
Posts Rated Helpful 1 Times
Quote:
Originally Posted by squeek. View Post
I haven't fixed the records to use a real Bayesian average yet, and I'm not totally sure I will. I lowered the constant of the Class Choice Percent records and am sort of regretting it; Wicked_Clown now has the offensive scout class choice record with 100% chosen in only ~21 rounds of offense played, making it a nearly unbeatable record if Wicked_Clown stops playing at this point. What I'm using the Bayesian average for is not to make it more average, it is to make outliers or unsustainable records not able to be the record holders in average categories. It's meant to make the records more consistent or meaningful, not more average.

You've only played the time equivalent of 5 rounds as engineer. The purpose of sorting by Bayesian average is to make it so even if you have 100% SG uptime in those 5 rounds, the system assumes that record is unsustainable. If the records were sorted purely by the real average and you held the record with 100% SG uptime and 5 pickups played, to keep that record you'd actually be encouraged to stop playing. The record couldn't realistically be improved, and the only thing that playing more would do is drive that record down. With the current system of Bayesian averaging, playing more actually increases the value of the record even if you aren't increasing your real average (in fact, you could even be decreasing it).

The current constant I'm using might be rewarding time played a bit too heavily, but I don't think by very much. Just so you know, the constant (C) I'm using right now is 7. The formula for each record is essentially:

(sum or count of relevant values) / (sum or count of relevant total + C)

I scale the C depending on the scale of the value/total as well, like for things that use sum of time played, I use 7*900 (900 seconds = 15 minutes, or one round).
Why aren't you summing an estimated average*C into the numerator
Isn't the formula bavg= (values + approxavg*C)/(C+numval)?
beees is offline   Reply With Quote