| If I understand this correctly, they are applying the Chess strategy of thinking n-moves ahead to Scrabble, where the opponent's tiles are hidden and ones own future tiles are mostly unknown. If you use all the tiles, on the upside you get +50 extra points. The downside is that you leave a lot more room for the opponent to score big since there are more letters on board after your move. Also, you have no control on your next 7 tiles and this can negatively affect your scores for the next 3-5 moves. The benefit to keeping a few good tiles for future moves comes at the cost of fewer points in the immediate round. The upside is your opponent has fewer open tiles to score big on. In general, the first strategy beats the second because of the bingo advantage, however their secret sauce is a list of 5 letter words that the opponent can't build on to score big points. So these inert words make the second strategy better for now. I haven't written any Scrabble AI but I wonder if most of them optimize for the immediate move or look further down the game? Unlike Chess, there are many more unknown variables per move in Scrabble - tiles you both might get, x-y position of n-tiles placed per round, scores of each tile placed, tile bonuses (double/triple letter/word) etc. If anything, the Nigerians' victory has shown that there is a lot more room for optimization in Scrabble and scoring the max per round may not be the globally optimal solution. |
Yep, Scrabble is a hidden-information, stochastic game (the opponent's future tiles are also randomly drawn).
Scrabble is probably more similar to poker and other card games in that respect. The key difference is that the distribution of letter tiles is skewed: the number of tiles is not the same for each letter [1]. This is different than, say, playing cards, where you get exactly 4 of each card (plus the, I think two, Jokers).
Which means it's easier to make a prediction for what's coming out of the opponent's bag, possibly (they're more likely to see the more common tiles).
>> they are applying the Chess strategy of thinking n-moves ahead
Well, that's a universal m-player game strategy, not specific to chess. Minimax and all that, ja? Did I misunderstand what you mean?
Except of course in this case the optimal strategy for Scrabble seems to be under dispute (long or short words best?) and so programming an AI player with minimax might not be that straightforward.
[1] https://en.wikipedia.org/wiki/Scrabble_letter_distributions