[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [computer-go] future KGS Computer Go Tournaments - two sections?



From: "Michal Bazynski" <bazik@xxxxxxxxxxxxxxxxx>
> hmm i do not understand where the problem understanding why it's wrong to
> have two same/similar programs comes from... to continue David's example:
>
> assume MFoG and GNU can beat every other program out there 100% of the
> time.
> assume MFoG can beat each GNU version 70% of the time.
> let's have 2 GNU's and one MFoG play in a tournament.
>
> what is MFoG's chance of winning? 70? nope, it has come down to 50%. 3
> copies of GNU's and its down to 35%. and all that time MFoG is clearly
> best - 70% winning chance or better against every other opponent.

(Firstly, I am not really convinced that go is a game of chance, so this
kind of math might not be correct ;-)

You are right, if MFoG is set against _any_ of the GnuGos winning. MFoG's
chances against a particular GnuGo copy are improving, because we also have
to account for the "inter-copy cannibalism".

I might have expressed myself poorly, the idea with this tournament would be
to look at GnuGo_1 and GnuGo_2 as different ones (the idea being that they
would have different tuning properties, or be run on different hardware, or
having improved L&D/ladder/wahtever modules). That's why I also proposed to
run this tournament anonymously: all programs will be then distinct.

If we group together all the GnuGos in a generic category, we are in fact
entering the realm of the second kind of tournament, where I also agreed
that multiple entries are unfair.

****
Let me try with an analogy. I'm not sure if it is very good, but hopefully
it is good enough: if go playing programs are like players, we might say
that their programmers/teams are like the Japanese go
schools/families/clans.

If the goal was to find the school with the best player, then it would be
unfair to let more than one competitor from one school to enter the finals.
Only the best player from each school need to meet eachother. This is my
case #2, or Nick's formal league.

If the goal is to find the overall best player, or rank all players against
each other, then of course all of them need to be in the competition. A
school with many players will have better chances to get more places at the
top compared to a solitary player, but that can't be helped. This is my #1,
or Nick's casual league.

We see here that there's a different perspective than for human go: the
"best player" league is what counts for humans, while the "best school"
league is what counts for computer-go. That is understandable, because the
programs have no egos or commercial interests, but the programmers do.

> what would make it fair is if MFoG could enter as many copies as GNU-based
> folks can.

yes, why not? Possibly tuned differently. It is a good way to try different
combinations of parameters, because one doesn't often have the chance to
play against many other programs while developing.

regards,
Vlad
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/