I like this model from an AI perspective -- it certainly sounds as if the output (probability confidence of the candidate play) could act as another input to the move heuristic, along with leave equity, score, etc.
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 On 9/6/07, Austin Nichols <[EMAIL PROTECTED]> wrote: > > Perhaps an even better approach is to allow Q to assign subjective > probabilities to each word of being good. Currently, it assigns a 1 to > every word in its list and a 0 to every word not there, but that is not how > a human player works (in fact, the subj prob may be arbitrarily close to 0 > or 1, but should never be exactly 0 or 1, even for ME and YOU and THE, > though at some point you get close to the subjective probability that the > Earth will continue rotating which is 1 within the limits of machine > precision). This would need a much longer word list with transpositions and > misspellings and some phonologically correct blatant garbage like VINXERIC* > or somesuch. Once you could do this, you would also want Q to estimate the > subjective probability that its opponent thinks a word is good, and with > some practice, the best Q opponent (who would still have subj prob of 0 or 1 > for every word) could realize potentially big gains in expectation against a > fallible human player by playing the most plausible phony. The weakest Q > player might get phoneys challenged on a regular basis, of course, in which > case it would need to get more circumspect about playing them. Make sense? > Hard to program? >
