Hi Darryl, This is a fascinating topic that I thought a bit about 10 years ago. Thanks for bringing it up.
I'm not an author of Quackle, but I would note that modeling a computer opponent will be much more accurate and effective than modeling a human opponent. If the objective is to improve performance against other programs, then the opponent modeling approach should be very effective. Is that the objective of Quackle, or is the objective to be more effective against human opponents than other programs? I am skeptical that the approach would help enough against the average human opponent to outweigh the additional risk of blundering. The same model would infer that a computer opponent does not have a blank (or S or whatever tile) when the player did not have play some superior word X would often not be valid for a human. The human might not know X, might not know the hook necessary to play X, might not see that X was a better play (e.g., was unduly scared of the opening that playing X created), or might just have overlooked X. The issue is that an invalid inference that our opponent does not have a particular tile like an S or blank in theri rack leave would cause us to undervalue the risk of making huge openings in our simulations when we have a good rack leave, which would lead to silly and unnecessary blunders when we are already big favorites to beat most any human opponent. Another interesting question is whether opponent modeling would be made available when computing the opponent's response to our move in simulations (i.e., does the opponent get to use a model of us when computing its response in simulation). Steven Gordon On 1/12/07, Darryl Francis <[EMAIL PROTECTED]> wrote: > > > > > > > There's a short piece in the new New Scientist journal (13 January) > entitled "Robot learns to play dirty". See: > http://tinyurl.com/yxw4uh > > Part of the article says: > "Adding in this 'opponent modelling' greatly improved the program's > game, allowing it to beat Quackle, one of the best conventional > Scrabble programs, by five points on average." > > Would the Quackle authors care to comment?
