ACBot uses my approach, which was machine learned from millions of game
results rather than from quackle's superleave.  They should probably be
"relearned" in light of QI.

That approach suffers from no concept of tile synergy other than the
duplication penalty, so no set of parameters could ever achieve a good
statistical fit to quackle's superleave.  For example, racks with Q and a U
will get undervalued or racks with Q and no U (and no I) will get overvalued
(or more likely both in order to achieve the closest statistical fit).

The most practical approximation to tile synergy for humans is intuition
(possibly with a little tuning to avoid overvaluing ING, for example).
Given that human intuition has to be applied on top of whatever tile values
one would memorize, anything like the numbers John suggested or any of the
other numbers that have been published are likely good enough to get in the
ballpark.

Steven Gordon


On 5/11/07, Amit Chakrabarti <[EMAIL PROTECTED]> wrote:

  * sapphirebrand2000 ([EMAIL PROTECTED] <sheppardco%40aol.com>) [070511
12:38]:
> I must refrain from a definite opinion, since I cannot devote the
> time necessary to studying this issue. But: perhaps you are better
> off using Basic, or Maven's values or something, along with a V/C
> balance adjustment.

One might be able to improve on this somewhat. The goal is to come up
with a simple model for approximating rack leave values that involves a
small amount of memorization (preprocessing phase) plus a small amount
of in-the-head arithmetic (query phase). I don't want to propose any
particular model, but it seems that the way to go about it would be to
learn the parameters of the model from quackle's data, basically
treating quackle's "superleave" values as the "correct" ones. Sounds
like a machine learning problem to me. Perhaps someone with solid
machine learning background (not me) can suggest a suitable
off-the-shelf algorithm to create a concise model.

For starters, it might be worthing learning the "correct" parameters for
an ACBot-style model. ACBot has values for each possible number of
occurrences of each tile, plus a matrix of V/C mix values.

-AC

Reply via email to