In terms of terminology I really like the names
Michael Reitz came up with:

Basic Game (37 rings, 5+7+9 marbles)
Standard Game (37 rings, 6+8+10 marbles)
Tournament Game (48 rings, 6+8+10 marbles)

My main interest is that creating a Zertz AI has been
on my "to-do" list which is fast becoming my "when I
retire" list at the rate I seem to be going and
greatly admire someone who has done more than think
about it!

Zertz fascinates me as a candidate for an AI because
while it has a huge fan out which makes naive tree
search impractical, it does have long combinations of
forced moves that in some sense suggest that with the
correct pruning algorithms I am almost certain a good
AI is possible! Unlike, for example, go in which the
fan out is huge and much depends on hard to quantify
notions of "good shape", it seems to me that if one
successively deepened lines with forced captures one
could get quite far.

That being said, even following all possible captures
may still be too big a search tree (i.e., a naive
approach would be to say follow all forced captures
until you reach an unforced move) and so it is
probably the case that one needs to add some sense of
intention.

When I get some more time I do intend to look over the
source code and may possibly be able to make some
suggestions. In any case this is a wonderful tool to
be able to concentrate on the AI aspects since you
have done the heavy lifting of move generation and
display not to mention network connectivity!

Cheers,

Lyman


--- Martin Trautmann <martintrautmann@...> wrote:

> Hi Lyman,
>
> thank you very much for your suggestions.
>
> Am Montag, den 03.09.2007, 17:39 -0700 schrieb Lyman
> Hurd:
> > I played with the tournament rules. After trying
> > several "standard" rules games I came to the
> > conclusion that they are fundamentally less
> > interesting because one can never be put into a
> > position where you have to win by a method other
> than
> > claiming white marbles. With 5 total white marbles
> > someone can always claim 3. In the tournament
> > version, the 6 available white marbles can be
> split
> > 3-3 (as they often are in advanced play) and one
> has
> > to alter ones strategy.
> >
> I just had the intention to implement the rules from
> the board game. My
> instructions explained it like that and just the
> additional rings are
> variants I found on the internet. I am not sure
> whether the names
> "standard" and "tournament" are a good choice since
> the name
> "tournament" is many times overloaded.
> What is your concrete suggestion?
>
> > The Zertz AI violates the fundamental principle of
> > opening strategy on the small board. There is a
> > forced win for the second player if the first
> player:
> >
> > 1) Plays the first marble on the edge of the
> board, or
> >
> > 2) removes a non-corner ruing in the first move.
> >
> > Lesson 1 was brought home to me in my first
> serious
> > game on Richard's PBEM server! Rule 2 was only
> > pointed out to me somewhat later and I was able to
> > verify.
> >
> > At a minimum I would suggest changing the AI so
> that
> > it never fell into one of these two traps (the
> > equivalent of a Fool's Mate in chess). There is a
> way
> > in these situations for the second player to force
> the
> > capture of two white marbles while giving up 1
> white,
> > 2 grey and 3 black and leaving an empty board.
> Then
> > no matter where the first player plays on the now
> > empty board, the second player can gain an
> additional
> > two whites.
> >
> > An interesting next step would be to make sure
> that
> > the AI knows how to take advantage of either of
> these
> > two mistakes! The first is easier to deal with.
> It
> > essentially entails isolating the marble into a
> corner
> > and I believe in no case involves allowing the
> first
> > player an unforced move. Some of the cases for
> rule 2
> > lead to situations where you have to allow an
> unforced
> > choice of captures even though both branches lead
> to
> > the desired result.
> >
> I guess the current AI is no match for a player in
> that league. My
> current AI just does breadth first search with
> slightly deeper sequences
> in case of knock out moves. In order to create a
> competitive AI for
> Zertz you have to make plans like "can I isolate an
> interesting amount
> of marbles" or "can I knock out many marbles at
> once". Plan making
> humans work in move depths which are totally out of
> reach of breadth
> first searches. The combination of set-moves with
> ring-remove-moves
> creates a huge space of possibilities.
>
> It would be interesting to create a competitive
> strategy but I am
> currently not about to add many challenging
> features. If you like to
> also help with implementations I surely could help
> you to get a grip on
> the code or to add an easier API for AI development.
>
> In case you like Holtz, what would be your major use
> case?
> I started to use it as network game with friends but
> later I used it
> more as analysis tool for PBM and Little Golem. For
> the analysis case I
> guess better AIs would not be the point but maybe a
> variant-tree of
> played moves like in ZF1. Would you as expert player
> seek for this
> feature or is ZF1 just fine?
>
> kind regards,
>
> Martin
>
>