From the short description, it looks like a project from Google about training a deep network to play the game. The resulting bot was quite strong (or at least, quite strong for not using Monte Carlo tree search, at least). Without any more details though, it's hard to tell.
Implementing a Go player using neural networks and reinforcement learning is really fun and taught me a lot about how said techniques work. I only covered the (much easier) 9x9 board, but would recommend it as an exercise to those interested :)
I've had this in my backburner for some time already (I'm actually an avid 9x9 player and have lots of games already available for an initial training, but probably not enough to get a decent player just out of a neural network)
Comments
From the short description, it looks like a project from Google about training a deep network to play the game. The resulting bot was quite strong (or at least, quite strong for not using Monte Carlo tree search, at least). Without any more details though, it's hard to tell.
Here is the paper for Google's DNN-based move evaluation in Go: http://www.cs.toronto.edu/~cmaddis/pubs/deepgo.pdf
Thanks, I wasn't sure where I had the reference.
Great! Thank you!
Implementing a Go player using neural networks and reinforcement learning is really fun and taught me a lot about how said techniques work. I only covered the (much easier) 9x9 board, but would recommend it as an exercise to those interested :)
I've had this in my backburner for some time already (I'm actually an avid 9x9 player and have lots of games already available for an initial training, but probably not enough to get a decent player just out of a neural network)
Do you have a link for the Google thing?