On Sun, 2006-03-26 at 20:17 -0500, Amit Chakrabarti wrote:
> * Jason E Katz-Brown ([EMAIL PROTECTED]) [060326 14:50]:
> > yeah we don't optimize the blank case by throwing out obviously shitty
> > plays. olaugh, we should.
>
> Could we discuss, on the list, some good optimisations to perform in
> this case? I've toyed with some for QAT (QAT = Quick Analysis Tool, a
> command-line-and-ASCII-based analyser I wrote some time back that will
> be public-domained Real Soon Now) but didn't settle on anything I
> really liked. It's not always correct to simply throw out plays with
> static value far less than the highest, though this may be correctable
> with better static eval such as that in Quackle. More importantly, it's
> not clear that "throwing out" some plays using a heuristic actually
> speeds up the Steve Gordon move generator. Any good solution should
> drastically reduce the number of pointers you follow in the GADDAG.
I agree, that'd be cool!
We don't use GADDAG right now; code is there to use it if one wants the
backend too but olaugh says it's broken, but probably in a
not-too-hard-to-fix way.
The kind of optimization I was referring to was how to not go thru the
possibly time-consuming process of leave board geometry evaluation for
obviously worthless plays. This is only a noticably big deal with
blanks, and seems pretty easy. Quackle does leave evaluation on a
bajillion plays that are like c(AT) for 2 points. In an endgame these
are worth looking at, but in midgame simulations it's not worth
computing a value with which to compare these lame plays.
One easy optimization we should make now without any thought at all is
to cache the computed value for rack leaves while looking at a given
position. (we could cache these eternally, but in the future we plan to
make leave evaluation have a basis in what's left in the bag.) That
itself might make simulation with blanks much faster. Since we don't
perform board geometry heuristic analysis, this might be all the
improvement we really need for now as it'd make getting our static eval
value trivially fast.
In other news, olaugh's getting really close to a perfect endgame
solver, I implemented optional bonus square labels and tile value
labels, and anand's bug is fixed.
cheerio
Jason
Yahoo! Groups Links
<*> To visit your group on the web, go to:
http://groups.yahoo.com/group/quackle/
<*> To unsubscribe from this group, send an email to:
[EMAIL PROTECTED]
<*> Your use of Yahoo! Groups is subject to:
http://docs.yahoo.com/info/terms/
