On Fri, Nov 20, 2009 at 10:21:25AM +0100, Thomas Lavergne wrote:
> On Thu, Nov 19, 2009 at 08:25:53PM -0800, Seth Pellegrino wrote:
> > Apologies all for replying to myself so many times, but Darren Cook
> > has been kind enough to help me clear the matter up. Essentially, I
> > had misread his paper. I was trying to compute the expensive features
> > Rémi uses for progressive widening as a move selector for playouts
> > rather than the stripped down, speedy features he actually uses.
> >
> > Thank you all for your help in increasing my understanding.
> >
> > >> Also, Tom, I was wondering if you could speak to adding patterns
> > >> larger than 3x3. A nice feature of 3x3 patterns is that they can be
> > >> jammed into two bytes (4 possibilities [black, white, vacant, off
> > >> board] for 8 positions), which facilitates the creation of a lookup
> > >> table. I've yet to figure out as fast a method for 4x4 patterns (or
> > >> larger), so I was wondering both how you've done it and how much it
> > >> helped.
>
> A thing to always get in mind is compute incrementaly, so for each
> features you have to check each time you play a stone, if this play
> chane features somewhere on the board.
For the liberties, of course that applies since they are used very often
and take little memory (if you are sensible about it), but also always
get in mind to avoid universal truths - not all features pay off to be
maintained incrementally. E.g. for my heavy playouts, I tried to
incrementally update features (mainly is-a-self-atari), but it turned
out the cache hits for all the stored information _and_ the work
involved in finding out what parts of the board I actually need to
recompute noticeably outweighted the benefit. Maybe if I optimized it
further, I would reduce the costs, but hardly much below the current
costs of re-computing the features around the play condidate on each
move.
--
Petr "Pasky" Baudis
A lot of people have my books on their bookshelves.
That's the problem, they need to read them. -- Don Knuth
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
