3
$\begingroup$

I've been thinking if machine learning can be used to play the game Scrabble. My knowledge is limited in the ML field, thus I've seeking some pointers :)

I want to know how could I possibly build a model that picks a move from all the given valid moves of the current game state, and then plays the move and wait for the delayed reward. The actions here aren't static actions, they are basically selecting move to maximize the final score.

Is there any way to encode the valid moves and then use a model to pick those moves?

I've also considered the genetic approach, but I think if I can represent my move with a set of features (score, consonantVowels ration, rack leave score, #blank tiles after the move, ...etc), training a neural network like this could take a long time.

Another training related question, is it feasible to run the training on a GPU given that I will be waiting for a response (the new game state) from the opponent (e.g. Quackle) after every action?

Thank you :)

$\endgroup$
  • $\begingroup$ Why do you want to use machine learning for this problem ? Isn't a simple dictionary more efficient ? I understand you want to use machine learning for delayed reward, but when I think about Scrabble, I don't really see it as a problem : since you will have random letter the next step, there is no point to "keep your letters". When I play Scrabble, I just try to maximize the score for the current turn, and there is very little point of trying to maximize score for next turns. $\endgroup$ – Astariul Nov 26 '18 at 0:54
  • 1
    $\begingroup$ I think there're some features that could describe each move, more than just its board score. Some of those features are the rack leave score, the number of blank tiles it uses compared to the score it's going to give me, the number of hot spots it opens, ... etc. Therefore I need to somehow tune the parameters for these features. $\endgroup$ – hemoali Nov 26 '18 at 21:13
1
$\begingroup$

Scrabble was solved using reinforcement learning achieving super-human performance. If you are interested, you can start reading about it here: https://en.wikipedia.org/wiki/Maven_(Scrabble)#cite_note-1

You can also read about it here: http://www0.cs.ucl.ac.uk/staff/d.silver/web/Teaching_files/games.pdf (there're several slides about Maven)

$\endgroup$
  • $\begingroup$ The term "solved" has a very specific meaning in the context of game AI, which definitely doesn't apply here. Even though it's already quite old, as far as I'm aware the PhD thesis about Maven is probably indeed still one of the more important source on AI for Scrabble. The full thesis can still be downloaded from this link: project.dke.maastrichtuniversity.nl/games/files/phd/… $\endgroup$ – Dennis Soemers Dec 21 '18 at 11:10
  • $\begingroup$ From what I've seen, your resources are pointing to Maven which doesn't involve reinforcement learning! $\endgroup$ – hemoali Dec 21 '18 at 13:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.