MLA 2.0 uses a percentage of moves at node level for levels >=6, based on a
measure of disorder/entropy and how strong a move has been evaluated at
previous depths.

When trying it a higher level in the tree, MLA pruned this always loses. I think
giving up moves according to a merit calculation criteria does work, but not
fixed or percentage limit. The horizon effect for captures is also an issue in
Abalone.

I agreee with you Peer about the tactical part. Even a single stone moved from
one field to another can be a key to a tactical combination.

I also think that dynamics (threats, sumitos, ejection threats) is what makes
the position change significantly. And you might need to calculate to important
depth for prisons (unless u have a pattern matching solution or a rough
evaluation concerning this aspect - i'm working on this at the moment). Prisons
look like so "game of go" ladder problems to me.

David.





Selon peer_sommerlund <peer.d.sommerlund@...>:

> --- In abalone_prog@yahoogroups.com, david.malek@f... wrote:
> >
> > I read the search algortihm, no hash, nothing special but pure
> alpha beta, so i
> > wonderred about speed. But the evaluation functions part needs more
> > investigation.
> >
> > In fact, I tried to select only a few moves in MLA, and i was
> proven wrong
> > against MLA without the pruning.
> >
> > So i'm looking forward to see a program play good by picking lony a
> few moves by
> > looking 2 moves ahead at node level at any level. This would really
> make me
> > question my tests.
>
> As I recall this "fixed-move-count" search was made popular by the
> famous "Chess" program that dominated computer chess 1970-1980, but
> modern chess programs uses recursive null move pruning which is
> tactically superior. I would expect the same to be true in Abalone.
>
> AbaPro does a 1-ply search before pruning leaves, but has no limit on
> the number of moves examined. Tax 2.0 apparently does a 1 or 2 ply
> search, and then examines only 6 moves at full depth.
>
> The greater the difference between prune-depth and search-depth the
> more tactical weakness I would expect. Thus at 8 ply the difference
> would be 6 ply for Tax .. leading to several tactical weaknesses. At
> 8 ply a recursive null move search does a 6 ply test, thus only 2 ply
> difference. The greater the search depth, the greater advantage I
> would expect with recursive null move pruning.
>
> Just for the record: The very first versions of Nacre also used a
> variant of "fixed-move-count", and my experience was also that it was
> better to do a less deep but full width search. The present version
> of Nacre uses a variant of recursive null-move search.
>
>
>
>
>
>
> Yahoo! Groups Links
>
>
>
>
>
>
>
>