RE: [computer-go] Learning : was Chess programs versus go programs

["David Fotland" <fotland@xxxxxxxxxxxxxxx>]

Searching is clearly the key to strong go play.  Many Faces has three
searchers in it.

A fast alpha-beta searcher for local tactics that decides if a block can be
captured or
get 5 liberties/two eyes to escape.

A full board alpha-beta searcher.

A best-first search life and death module.

Neither alpha-beta searcher uses iterative deepening since the search depths
are so irregular.

The tactician looks typically 5 ply and up to 80 ply deep, with a budget of
a few hundred nodes
per search.  It has a transposition table, and it remembers the best move at
the root from move
to move, along with a list of board points that invalidate the search, to
reduce re-searching time.

The full board search looks 1 to 30 ply deep, with a budget of a few hundred
nodes.  The tree near
the root is taken from patterns or joseki library, and can be 1 to 20 or so
ply deep.  Then comes
a local quiescence search, then a global quiescence search.

I don't use null move because the evaluation and move generators understand
threats already.

The best first search is similar to PN search except that it uses
probability of success from
the move generator rather than number of generated moves to pick a node to
expand.

Most of my time goes into tuning the move generators.

Regards,

David

> -----Original Message-----
> From: computer-go-bounces@xxxxxxxxxxxxxxxxx 
> [mailto:computer-go-bounces@xxxxxxxxxxxxxxxxx] On Behalf Of Matt Gokey
> Sent: Tuesday, December 07, 2004 11:10 PM
> To: computer-go
> Subject: Re: [computer-go] Learning : was Chess programs 
> versus go programs
> 
> 
> Vincent, I don't necessarily disagree with you (about 
> search).  As you recall in my message, I summarized my point 
> like this:
> 
> >> There is little question in my mind that searching is a key to 
> >> computer-go just as it was in chess.  It may not be the 
> same kind of 
> >> search (muti-level for example), and the evaluations 
> (including life 
> >> and death) may not be as easy but I think its the most promising 
> >> direction to pursue in combination with pattern based concepts for 
> >> move generation and help in evaluation.
> 
> So, the search technique is primary, of course paired with 
> other things including solid evaluation.  Pattern or rule 
> based module components could be hand coded, or auto 
> generated/learned, or even coded by a 1000 monkeys as long as 
> they work reasonably well.  But I've done enough hand tuning 
> to know that intuition after a few iterations is as often 
> wrong as it is right.  There is a lot of trial and error and 
> guessing, albeit educated guessing. Each method has its pros 
> and cons.  I'm well aware that you don't see value in 
> automated learning or pattern harvesting.  And I certainly 
> don't expect any learning system to "beat" human learning 
> capacity.  That's not the goal or the expectation, however.   
> There is something appealing about the objective and 
> statistical nature of the pattern harvesting technique that 
> Frank has developed over the ad-hoc and "expert" tuned 
> pattern/rule sets commonly used.  I would not discount it as 
> a useful part of a go-playing program.  As Frank has said 
> it's designed to be a better Joseki/Fuseki database, not the 
> holy grail.  There is nothing preventing combining it with 
> other modules using other techniques for handling some 
> tactical issues, local search, ladders, ko fights, life/death 
> analysis, etc.
> 
> You seem to have a tendency to try to frame things as black 
> and white only, then attack the extreme case, when no-one is 
> really talking about the extreme case.
> 
> Regards,
> 
> Matt
> 
> 
> _______________________________________________
> computer-go mailing list
> computer-go@xxxxxxxxxxxxxxxxx 
> http://www.computer-go.org/mailman/listinfo/computer-go/
> 


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