4 ms·

Not entirely on topic, but: I would really like to see how DeepMind or another sophisticated task-specific AI would approach high level scrabble. At the highe

by airza 6y ago

Not entirely on topic, but: I would really like to see how DeepMind or another sophisticated task-specific AI would approach high level scrabble. At the highest level, constant rearrangement of slamming 7 letter words onto the board seems to dominate all other concerns, seeming to dominate the strategy of trying to close the board up and choke the opponent out. It would be interesting to see what other strategies emerge, or if a dominant AI simply has great statistical foresight in how to minimize the time between bingos!

cdelsolar 6y ago
I'm working on an AI and have some notes here: https://domino14.github.io/macondo/ https://domino14.github.io/macondo/ The state of the art AIs (and this one) use truncated Monte Carlo search to determine best plays. They can still be beaten by top humans around 50% of the time. We will be dedicating a lot of time next year to adding machine learning, inferencing (i.e. range finding ala poker), and other approaches to try to build the best AI in the world.
- dmurray 6y ago
  Well, this goes against what I said above, and you clearly know the state of play better than I do. Have you got a source for top humans beating top AIs 50% of the time? I found this where the top answer mentions a 500-game series between Kenji Matsumoto and Quackle that sounds pretty even: https://www.quora.com/Is-there-a-Scrabble-computer-program-that-beats-the-best-in-the-world-much-like-Deep-Blue-beat-the-best-Chess-player-in-the-world https://www.quora.com/Is-there-a-Scrabble-computer-program-t...
  - cdelsolar 6y ago
    yeah, I know Kenji - he's a top player and can beat the computer about 50% of the time. Will Anderson, another top player, is doing a 100 game series with Quackle on Twitch and they're roughly even after 20-ish games IIRC. I'm not as good as them but I can still beat it probably 35-40% of the time.
dmurray 6y ago
There's been world champion level Scrabble AI since around 2002 [0]. There hasn't been anything groundbreaking on it since: not because nobody has thought of it as a research problem, but because, unlike chess, improvements to the algorithm haven't led to great improvements in performance. This would suggest that Maven is pretty close to perfect: the game just isn't that deep, and the randomness of the tiles drawn dominates any small performance gain. For human players a large part of the game is in knowing the word list: even world champions spend hours poring over different permutations of that, which a computer can access instantly. The strategic part of the game seems interesting, but can be almost completely accounted for with some simple heuristics and the knowledge of which tiles remain. [0] https://www.sciencedirect.com/science/article/pii/S0004370201001667 https://www.sciencedirect.com/science/article/pii/S000437020...

by airza 6y ago

cdelsolar 6y ago

I'm working on an AI and have some notes here: https://domino14.github.io/macondo/ https://domino14.github.io/macondo/ The state of the art AIs (and this one) use truncated Monte Carlo search to determine best plays. They can still be beaten by top humans around 50% of the time. We will be dedicating a lot of time next year to adding machine learning, inferencing (i.e. range finding ala poker), and other approaches to try to build the best AI in the world.

dmurray 6y ago
Well, this goes against what I said above, and you clearly know the state of play better than I do. Have you got a source for top humans beating top AIs 50% of the time? I found this where the top answer mentions a 500-game series between Kenji Matsumoto and Quackle that sounds pretty even: https://www.quora.com/Is-there-a-Scrabble-computer-program-that-beats-the-best-in-the-world-much-like-Deep-Blue-beat-the-best-Chess-player-in-the-world https://www.quora.com/Is-there-a-Scrabble-computer-program-t...
- cdelsolar 6y ago
  yeah, I know Kenji - he's a top player and can beat the computer about 50% of the time. Will Anderson, another top player, is doing a 100 game series with Quackle on Twitch and they're roughly even after 20-ish games IIRC. I'm not as good as them but I can still beat it probably 35-40% of the time.

dmurray 6y ago

There's been world champion level Scrabble AI since around 2002 [0]. There hasn't been anything groundbreaking on it since: not because nobody has thought of it as a research problem, but because, unlike chess, improvements to the algorithm haven't led to great improvements in performance. This would suggest that Maven is pretty close to perfect: the game just isn't that deep, and the randomness of the tiles drawn dominates any small performance gain. For human players a large part of the game is in knowing the word list: even world champions spend hours poring over different permutations of that, which a computer can access instantly. The strategic part of the game seems interesting, but can be almost completely accounted for with some simple heuristics and the knowledge of which tiles remain. [0] https://www.sciencedirect.com/science/article/pii/S0004370201001667 https://www.sciencedirect.com/science/article/pii/S000437020...