solutions to solved puzzles as training data for machine learning model

Started by jlc206

jlc206 Lv 1

I know there are videos on youtube of solved puzzles, but does there exist some open database of the moves that were taken to get to the solution of a solved puzzle? I'm thinking along the lines of training a machine learning model to predict the next move based on the previous moves and current state of the puzzle (basically learn how players play)

Thanks in advance!

jeff101 Lv 1

I am interested in this too. Below are some related posts:
a recent paper about a puzzle done in Fall 2015

a movie from the above paper showing how the structure evolved during the puzzle
another paper detailing the moves done in the above puzzle & movie
"Using Neural Networks to learn from Foldit players"
has links to some youtube videos showing how the
structure evolved during certain Foldit puzzles

TomTaylor Lv 1

Your suggestion is why foldit exists. Scientists are trying to determine how humans solve folding puzzles and eventually use computers to simulate those techniques. It's a very difficult undertaking and they are still in the early stages.

rmoretti Staff Lv 1

This is certainly something we're interested in.

While not involving machine learning, there has already been a paper about analyzing what Foldit players have done with recipes ("Algorithm discovery by protein folding game players"

The Foldit client does indeed collect limited information about the sorts of moves players make (though this is not an open database), and there's been some thought about using this information in an attempt to help divine the "special something" that human intuition provides in the Foldit process.

free_radical, in particular, has been in contact with several groups doing advanced machine learning with that in mind. The feedback he's gotten so far is that the type of information we currently have is not exactly what those groups would be looking for to do this sort of machine learning.

That said, we're still interested in further algorithm discovery studies, and machine learning will likely play a role in this. If we come up with a way to effectively use the current move information, or if we come up with a way to usefully augment the move information, we'll keep you posted.