New Giraffe

[matthewlai]

Matthew,

A while back you specualted about using NN in the search as well as having a NN evaluation function.

Is this still a realistic goal and would it help you if you had access to more cores?

Carl

It's already in the last version (called move evaluator network). I just haven't ported it to this new version yet, and I suspect it may be responsible for the difference in playing style.

[Dann Corbit]

I do not know much about torch.
Does your current code take advantage of GPU cards?
I would be especially interested to see if it can run on something like Jetson, with integrated memory:
http://www.nvidia.com/object/jetson-tk1 ... v-kit.html

It is a tiny GPU, and the CPU is wimpy, but there is no copying to and from RAM because GPU and CPU can share the same fast memory.

[Dann Corbit]

I guess that this is the torch that you used, is that right:
http://torch.ch/

[CheckersGuy]

Matthew,

A while back you specualted about using NN in the search as well as having a NN evaluation function.

Is this still a realistic goal and would it help you if you had access to more cores?

Carl

It's already in the last version (called move evaluator network). I just haven't ported it to this new version yet, and I suspect it may be responsible for the difference in playing style.

Is this search-Neural-Network used for moveOrdeirng or does it pick some of the possible moves and disregards everything else ?

[matthewlai]

I do not know much about torch.
Does your current code take advantage of GPU cards?
I would be especially interested to see if it can run on something like Jetson, with integrated memory:
http://www.nvidia.com/object/jetson-tk1 ... v-kit.html

It is a tiny GPU, and the CPU is wimpy, but there is no copying to and from RAM because GPU and CPU can share the same fast memory.

It does not use GPUs currently. At play time the CPU-GPU latency is higher than the time it would take to just evaluate on the CPU.

A heterogeneous architecture may change that, but I don't have one to play with (and don't really have time at the moment even if I do).

Yes, torch.ch is the right Torch.

[matthewlai]

Matthew,

A while back you specualted about using NN in the search as well as having a NN evaluation function.

Is this still a realistic goal and would it help you if you had access to more cores?

Carl

It's already in the last version (called move evaluator network). I just haven't ported it to this new version yet, and I suspect it may be responsible for the difference in playing style.

Is this search-Neural-Network used for moveOrdeirng or does it pick some of the possible moves and disregards everything else ?

It uses the move evaluator network to assign probabilities in a probability-based search. See the paper (chapter 5) for more details: https://arxiv.org/pdf/1509.01549v2.pdf

[Sylwy]

http://talkchess.com/forum/viewtopic.ph ... 68&t=50060

[Graham Banks]

Hi Matthew,

sent you a message regarding a possible problem.

Graham.

[matthewlai]

If it finds eval.net and eval.t7 in the same directory, which one will it use?

Also, does it not need a meval file as well?

eval.t7. It doesn't look for or care about *.net anymore. The old format is not supported anymore.

Move evaluator (will be meval.t7) has not been ported to this new version, yet. That will come later.

[matthewlai]

Move evaluator (will be meval.t7) has not been ported to this new version, yet. That will come later.

Will it impove playing strength considerably? I ask this because my first test results are rather disappointing, showing no improvement at all over version 2015.09.08.

I am hoping yes. It seems to be quite a bit stronger at very fast games (0.5s per move), but that difference disappears at longer time controls.