> Regarding 9x9, I believe Alvaro Begue has explored this idea in a way > which perhaps would work better in a go engine. He used pachi to generate a > database of games by playing against itself and then trained a model in a > similar fashion to what I did. I'm not sure about the results of his > experiments. If someone can point me to a large database of 9x9 games it > would be easy to edit my code to do that. >
My experience is the same: My CNN was a very poor judge of life and death. Part of the problem is that I couldn't get Pachi to behave exactly the way I wanted (play to maximize score; play to the bitter end, assuming everything left after two passes is considered alive). But perhaps there is some deeper problem, or we are just missing an important twist to make the technique work. Since those initial 9x9 experiments I have worked on the much easier problem of coming up with a probability distribution for the next move, to make sure I wasn't doing anything really wrong with the neural network. It seems to be working well enough (48% accuracy with a very limited set of inputs), so I think I'll switch back to trying to predict ownership and score. Álvaro. On Tue, Jan 12, 2016 at 6:10 PM, Justin .Gilmer <jmgil...@gmail.com> wrote: > Quick question: When using this mailing list, how to I explicately reply > to a thread, so far I've just been editing the subject and sending it to > computer-go@computer-go.org. > > Regarding use in a MTCS engine, I strongly suspect it would perform poorly > in its current form. It is quite poor at life and death, especially if you > give it situations very different from the training set. One issue with the > method of training was I only used games which were played until the end > (i.e. didn't end in resignation), as a result the model is extremely biased > that large groups of stones live simply because games not ending in > resignation tend to be close and not have large groups die. > > Depending on how hard it would be to integrate into a MCTS I could try it. > My hope was that a well trained evaluator could allow for alpha beta > pruning to be competitive with MCTS, interested to hear the groups thoughts > on this. > > Regarding 9x9, I believe Alvaro Begue has explored this idea in a way > which perhaps would work better in a go engine. He used pachi to generate a > database of games by playing against itself and then trained a model in a > similar fashion to what I did. I'm not sure about the results of his > experiments. If someone can point me to a large database of 9x9 games it > would be easy to edit my code to do that. > > > -Justin > > > > _______________________________________________ > Computer-go mailing list > Computer-go@computer-go.org > http://computer-go.org/mailman/listinfo/computer-go >
_______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
