Modeling likely challenge responses using an imperfect probabilistic lexicon
model of Q's opponent is definitely more complex, but with a clear advantage
against a lexicon-limited player... e.g. holding ACILLVY the expected value
of CAVILLY for (say) 87 or zero (dep on whether it gets challenged) static
value versus CAVIL for 25, or slightly more if it gets challenged, depends
crucially on the probabilities you assign to whether your opponent will
challenge.

I thought the usefulness of the modeling for making Q stronger might make
the strategy appealing to Q's developers, and the possibility of using an
imperfect model of the lexicon (used by Q for inference about fallible
players' responses) as Q's own lexicon for making Q play at a lower level is
just a side benefit.  A big side benefit.

Note that Q against Q might also produce phoney plays if Q doesn't know it's
playing itself, and simulations might produce an optimal mixed strategy...

To make an expanded lexicon with phoneys, you can start with some manual
entry of common misspellings and then automatically generate using
transposed vowels and noun+ED or verb+LY constructions and the like.  The
construction of plausible-looking phoneys is an independent exercise, of
course, from the programming task of allowing an imperfect probabilistic
lexicon model.  However, a very large list of phoneys could be whittled down
to a better list by letting Q play real opponents and learn which phoney it
could sneak by real players.

On 9/6/07, John Van Pelt <[EMAIL PROTECTED]> wrote:
>
> It would probably be more complex to include in the sims the probability
> computation (per move) that opponent might challenge -- figure the
> probability confidence that the given word carries a given probability
> confidence in the opponent's lexicon?
>
> The real question is how to seed a starter lexicon with data like this?
>
> jvp
>
>
>

Reply via email to