Re: [Bug-gnubg] A new variance reduction technique
From:
Massimiliano Maini
Subject:
Re: [Bug-gnubg] A new variance reduction technique
Date:
Tue, 31 Mar 2009 14:26:04 +0200
address@hidden wrote on 31/03/2009 14:10:30:
> 2009/3/30 Massimiliano Maini <address@hidden>:
>
> > A few comments (assuming my understanding is right):
> >
> > 1. Instead of going for the same rank, you could (at negliectible
cost)
> > select
> > the roll for play B that gives the closest luck to the one obtained
for play
> > A.
> > Should be more efficent than just taking the roll with the same
rank.
>
> This will introduce errors since not all rolls in B will be chosen
> with the same probability.
Well, you have the same problem when you chose according
to the rank: you pick randomly in A, but then non-randomly in B.
It's the goal. Or am I missing something ?
> All in all an interesting idea, but complex and error prone. I am
a
> little afraid that we might throw the baby out with the bathwater
-
> since the current rollout code has some pretty complex improvements
as
> well (Quasi Random Dice AKA Stratification).
I thought about how this mixes with stratification
(the naming 'quasi random dice" sucks big time, IMHO). I think Bob's idea
should only be applied to rolls subsequent to stratification. E.g.: 1296
trials = 36^2, then stratify 1st and 2nd roll, use Bobs' method 3rd roll
on. I also think that Variance Reduction (Dahl's) should
not be applied on perfectly stratified rolls since it can just introduce
errors: average luck over the 36 (1296) rolls is zero.
Anyway, I'm not convinced Bob's idea will improve
by much: it's not even sure it will improve at all (compared to Var. Red.).