On Fri, Jan 23, 2009 at 7:29 PM, Graham Toal <[email protected]> wrote: > On Fri, Jan 23, 2009 at 6:06 PM, Eugene Deon <[email protected]> wrote: >> I've been trying to tune my leave estimation strategy by solving for the >> exact values of my commonly considered estimation variables as to minimize >> the sum of squared errors between my strategy estimates and all the leaves >> in Quackle's "superleaves" file. >> >> So far my variables include only: >> >> -single letter values >> >> -double/triple/quad-letter penalties >> >> -vowel/consonant imbalance penalties >> >> -bonuses for number of tiles in CANISTER >> >> -a couple of letter pair values (QU, YY, IY, FF, ING) > > Why all the special cases? The value of a leave is how much it > contributes to the next play. > > So... enumerate every possible play that can be made next with the > unseen tiles and the leave. (This sounds like a lot of computation > but it isn't) > > For each word, calculate the probability of drawing the tiles you need > to make the play given the tiles you are holding (ie the leave). (I > have the code that enumerates that probability function) > > Do this for other leaves, and compare. Special cases like letter pair > values etc fall out in the wash; in fact, it works for larger leaves > the same way (and faster, as the play choices are more limited) > > Graham
Graham, This approach would effectively turn an N ply simulation into an N+1 ply simulation, except that the last ply would be exhaustive instead of random. I cannot believe that could be nearly as efficient as a static rack evaluation at the leave nodes of a simulation. Steve
