It could just be that the Scrabble AI is not sufficiently advanced to explore the entirety of the decision tree. For all board states X, there are many deep trees that increase in complexity at boards X+1, X+2, etc (though it's arguable that it decreases in complexity in the late game, or can a game can be played "closed" to reduce the chances of bingos).
It may not be purely metagaming that causes the difference between AI and humans; the algorithm probably plays a far-from-perfect game.
The other advantage a computer (probably?) has over a human, is the analysis of probabilities of words being available given whats known about the word play of the opponent given their current (unknown) tiles, and tiles already played.
That is, I suspect it's possible, but difficult for a human player to surmise the probability of tiles hidden in an opponents pieces based on the words they've recently played (or not played) - but being able to do so, based on a count of which tiles have already been played and are visible on the table, is probably more challenging - but straightforward for a computer.
Are the games that players have played publicly available? While I presume that players at this level have memorized all 270K or so words, they might have tendencies to favor some words versus others, in particular scenarios - that's another advantage that a computer would have - adjusting it's play based on a players history.
>That is, I suspect it's possible, but difficult for a human player to surmise the probability of tiles hidden in an opponents pieces based on the words they've recently played (or not played) - but being able to do so, based on a count of which tiles have already been played and are visible on the table, is probably more challenging - but straightforward for a computer.
I have played a very small amount of competitive Scrabble and a few of my friends are playing in the National tournament that this article references. What you describe is utterly standard by even mediocre tournament players.
Tournament Scrabble players do a lot more than memorize good lists of 2s and 3s. It is truly impressive how deep the game can be.
The article claims there are just shy of 200,000 scrabble words.
Why would it be so complex? The key is to use all 7 of your letters. The subset of valid words containing all 7 letters you hold shouldn't be that large, surely the board can be exhaustively searched to fit them?
You don't just want good words, you want to position yourself so your next move is also a good word. So you don't just have to iterate over the current board, you have to iterate over all possible upcoming boards.
But it's even worse than that! You want to set up your opponent for a bad move, and vice versa, so you're trying to infer each others' tiles from past moves and bluffing starts to matter. So you also have to reason about counterfactual games where your opponent would have done the same thing, meaning you have to care about the entire game tree instead of just the sub-tree you're in...
In the general case, solving games involving partial information requires super-exponential time. It's no picnic. (Of course in practice we use approximations, and those can work quite well.)
Good point regarding setting up for double/triple word scores, however I doubt there is much soul reading or bluffing in scrabble - I think the cost of making a sub-optimal move is too great, you'd need to be very sure that it hurt your opponent even more than yourself.
Also although the remaining letters are known, your selection is random. I don't know how you could possibly choose between words to maximize your next move when you don't know what letters you will get?
> I don't know how you could possibly choose between words to maximize your next move when you don't know what letters you will get?
Your selection is random amongst a determined pool of tiles. You can easily maximize your move even if the move is not 100% deterministic. The fact that there is a random factor does not mean strategy is useless.
This is trivially true in a lot of cases, but sticking with games: The dealer's concealed (hole) card in Blackjack is not known precisely but by counting cards, you can determine the best play even if the effect is small.
You don't have to use all your letters for the same word, and there are strategic considerations regarding which spots you place your words, since that might enable the other player to use the score multiplier tiles for their next move.
IIRC 15k-20k is usual for monolingual adult speakers of English. Since different languages have different total vocabulary size -- English's is particularly high -- it probably varies quite a bit across languages. I believe bilinguals tend to have a smaller vocabulary in each language, but larger if both are counted.
You can estimate your English vocabulary at http://testyourvocab.com/ (BTW, if anybody knows a site that does a similar test for other languages -- particularly German -- I'd be interested).
It may not be purely metagaming that causes the difference between AI and humans; the algorithm probably plays a far-from-perfect game.