[Computer-go] MCTS + Neural Networks?

Lucas, Simon M sml at essex.ac.uk
Wed May 1 23:58:58 PDT 2013

A recent comparison of move prediction systems for Go
can be found here (IEEE CIG 2012 proceedings)


Simon Lucas

From: computer-go-bounces at dvandva.org [mailto:computer-go-bounces at dvandva.org] On Behalf Of Steven Clark
Sent: 02 May 2013 03:10
To: computer-go at dvandva.org
Subject: Re: [Computer-go] MCTS + Neural Networks?

Thanks for the link! Looks like a good paper -- I will read it more carefully shortly.
Ignoring computational speed for a moment, is it a reasonable assumption that an algorithm that plays a NN-proposed tactical move 50% of the time, and a random move 50% of the time, should outperform an algorithm that plays a random move 100% of the time?
So it's just a case of how many playouts do we lose by employing the NN (GPUs to the rescue?). For reference, I was using 25 input nodes, 25 output nodes, ~50 hidden nodes.
I guess ultimately it comes down to "make a bot and prove it" :)


On Wed, May 1, 2013 at 9:50 PM, George Dahl <gdahl at cs.toronto.edu<mailto:gdahl at cs.toronto.edu>> wrote:
I don't know if neural nets that predict moves have been helpful in any strong bots, but predicting expert moves with neural nets is certainly old news. See http://www.cs.utoronto.ca/~ilya/pubs/2008/go_paper.pdf

There might be a place for artificial neural nets in a strong Go playing program, but it is an open question on how to incorporate neural nets well. Software like neurgo used a lot of expert features along with a neural net for global position evaluation and I tried (with very little success) to predict ownership of points on the board using a neural net.

It is very hard to get neural nets to help a standard MCTS bot a lot because the neural net needs to be good at whatever it is supposed to be doing and still probably very fast to be useful.

- George

On Wed, May 1, 2013 at 9:42 PM, Steven Clark <steven.p.clark at gmail.com<mailto:steven.p.clark at gmail.com>> wrote:
Hello all-

Has anyone successfully used neural nets to help guide MC playouts?
Has anyone used NN to learn patterns larger than 3x3?

I'm working on a grad-school project, and discovered a few interesting things.
After analyzing 10,000+ high-dan games from KGS, I find that more than 50% of the time, moves are played within a 5x5 window centered at the opponent's previous move (call this a "tactical" move, vs a strategic move).

I used the FANN library to learn these 5x5 patterns, and found that the NN could predict tactical moves with ~27% accuracy (and with a >50% chance that the answer would be in the top 3 moves proposed by the NN).

Is this old news? Are neural nets just too slow to be helpful to MC (reduce the playout rate too much?)

Thoughts welcome. I will be up late finishing the report since it is due tomorrow ;)


Computer-go mailing list
Computer-go at dvandva.org<mailto:Computer-go at dvandva.org>

Computer-go mailing list
Computer-go at dvandva.org<mailto:Computer-go at dvandva.org>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://computer-go.org/pipermail/computer-go/attachments/20130502/fc37438a/attachment.html>

More information about the Computer-go mailing list