Snowie v. JellyFish
EdmondT
These independent random number generators gave me a great idea of how to test
the two programs for quality. Have two matches between the two programs,
playing different sides of the same randomly generated numbers. See which
program does better on both.
Wouldn't this prove, beyond a doubt, which was better? (Assuming a large
number of games was played?)
Chuck Bower
In article <199806181731...@ladder01.news.aol.com>,
EdmondT <edm...@aol.com> wrote:
Please define "large". I guess the real question is: "do duplicate
dice reduce the variance compared to random dice, and if so, by how much?"
I believe Fredrik has investigated this in regard to Jellyfish rollouts.
Maybe he will toss in this answer while replying to the earlier post: ("when
is the next version of JF to be released?) In fact, he probably already
answered this on this newsgroup. DejaNews, anyone?
As has been mentioned many times here before, there was a book pubbed
in the late 70's by Barclay Cooke and Rene Orlean which annotated part of
a team match--US=Barclay Cooke and his son Walter, UK=Joe Dwek and Philip
Martyn. That match was performed with duplicate dice. The authors'
conclusion was that duplicate dice don't buy you anything in terms of
eliminating the luck factor and therefore are not worth the added pain.
The reason is that one change in play (one room being different that the
other) led to a divergence at the two tables. Most common example: 66
is great for running and building, but not so pleasant most of the times
you are riding the pine. I don't know of any high level competition
since that match (which was played in the early 70's, I think). I guess
the rest of the BG world agreed with their assessment.
Chuck
bo...@bigbang.astro.indiana.edu
c_ray on FIBS
Fredrik Dahl
Chuck Bower <bo...@bigbang.astro.indiana.edu> skrev i artikkelen
<6mbmi6$gtk$1...@jetsam.uits.indiana.edu>...
> In article <199806181731...@ladder01.news.aol.com>,
> EdmondT <edm...@aol.com> wrote:
>
> >These independent random number generators gave me a great idea of how
to test
> >the two programs for quality. Have two matches between the two
programs,
> >playing different sides of the same randomly generated numbers. See
which
> >program does better on both.
> >
> >Wouldn't this prove, beyond a doubt, which was better? (Assuming a
large
> >number of games was played?)
>
>
> Please define "large". I guess the real question is: "do duplicate
> dice reduce the variance compared to random dice, and if so, by how
much?"
> I believe Fredrik has investigated this in regard to Jellyfish rollouts.
> Maybe he will toss in this answer while replying to the earlier post:
("when
> is the next version of JF to be released?) In fact, he probably already
> answered this on this newsgroup. DejaNews, anyone?
>
:-)
Fact is that this variance reduction technique helps very little;
play one roll differently in the opening, and the relative
value of different rolls for the rest of the game is altered dramatically.
If you play two parallel games with opening roll of 43 by 13/10 24/20 and
13/9 24/21, you will see that the games diverge extremely fast.
A 66 for player one on roll 6 can easily be a disaster dancing roll
in one game, and a perfecto in the other.
About JF we are planning a 3.5 release soon.
It will not have significant new featurs, but the UI is
improved. More on this later.
Fredrik Dahl.
David Montgomery
In article <6mbmi6$gtk$1...@jetsam.uits.indiana.edu> bo...@bigbang.astro.indiana.edu (Chuck Bower) writes:
>I guess the real question is: "do duplicate dice reduce the
>variance compared to random dice, and if so, by how much?"
>I believe Fredrik has investigated this in regard to Jellyfish rollouts.
I investigated this with JF rollouts, and my results were that
duplicate dice didn't help at all.
Fredrik told me that "these techniques are far less effective than
you might think," but he was surprised that I didn't find any benefit.
He felt that the benefit for races and priming games should be pretty
big.
My experiments were pretty crude, but I don't think duplicate dice
are worth much for most positions.
However, there is one situation where they have a lot more promise.
This is when you are going to roll out the *same* play at different
match scores. As Chuck wrote, the problem with using duplicate dice
for play vs play rollouts is that the positions quickly diverge, and
so what is lucky in one position may be unlucky in the other.
However, if we are rolling out the same position at two different match
scores, we might expect that the two rollouts would diverge much more
slowly, since they start out the same, and very often the correct play
at one score will also be correct at the other.
David Montgomery
mo...@cs.umd.edu
monty on FIBS