Re: [Computer-go] AlphaGo Zero
1) There is no such thing and I do doubt if it ever will exist. Even humans fail elaborate why they know certain things 2) If we are talking about new one. Very few people seen it playing so I guess we lack the data. For the old we know it made errors, dunno if analysis points why. Neural nets tend to be black boxes 3) Would it be a bad thing? All thing considered, not just human point of view 2017-10-20 7:06 GMT+03:00 Robert Jasiek : > So there is a superstrong neural net. > > 1) Where is the semantic translation of the neural net to human theory > knowledge? > > 2) Where is the analysis of the neural net's errors in decision-making? > > 3) Where is the world-wide discussion preventing a combination of AI and > (nano-)robots, which self-replicate or permanently ensure energy access, > from causing extinction of mankind? > > -- > robert jasiek > > ___ > 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
Re: [Computer-go] agz -- meditations
Cost reduction in IC has reached or is reaching its limits. Intels 5n techk is not really a 5n and 5n is not really reachable. Not at least without some seriously new physics and even then there will be hard limits like quantum un--certainty. This particular chip may get cheaper if it is ever done in amounts but it is not guaranteed to get lot cheapaer 2017-10-20 6:24 GMT+03:00 Robert Jasiek : > On 19.10.2017 20:13, Richard Lorentz wrote: > >> Silver said "algorithms matter much more than ... computing". >> Hassabis estimated they used US$25 million of hardware. >> > > Today, it seems 4 TPU cost US$25 million. In 5 or 10 years, every computer > might have its 4-TPU-chip costing $250, if not $25. At least, I hope. > > -- > robert jasiek > > ___ > 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
Re: [Computer-go] AlphaGo Zero
So there is a superstrong neural net. 1) Where is the semantic translation of the neural net to human theory knowledge? 2) Where is the analysis of the neural net's errors in decision-making? 3) Where is the world-wide discussion preventing a combination of AI and (nano-)robots, which self-replicate or permanently ensure energy access, from causing extinction of mankind? -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] agz -- meditations
On 19.10.2017 20:13, Richard Lorentz wrote: Silver said "algorithms matter much more than ... computing". Hassabis estimated they used US$25 million of hardware. Today, it seems 4 TPU cost US$25 million. In 5 or 10 years, every computer might have its 4-TPU-chip costing $250, if not $25. At least, I hope. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo Zero
Yes, residual networks are awesome! I learned about them at ICML 2016 ( http://kaiminghe.com/icml16tutorial/index.html). Kaiming He's exposition was fantastically clear. I used them in my own attempts at training neural networks for move prediction. It's fairly easy to train something with 20 layers with residual networks, even without using batch normalization. With batch normalization apparently you can get to hundreds of layers without problems, and the models do perform better on the test data for vision tasks. But I didn't implement that part, and the additional computational cost probably makes this not worth it for go. Álvaro. On Thu, Oct 19, 2017 at 8:51 PM, Brian Sheppard via Computer-go < computer-go@computer-go.org> wrote: > So I am reading that residual networks are simply better than normal > convolutional networks. There is a detailed write-up here: > https://blog.waya.ai/deep-residual-learning-9610bb62c355 > > Summary: the residual network has a fixed connection that adds (with no > scaling) the output of the previous level to the output of the current > level. The point is that once some layer learns a concept, that concept is > immediately available to all downstream layers, without need for learning > how to propagate the value through a complicated network design. These > connections also provide a fast pathway for tuning deeper layers. > > -Original Message- > From: Computer-go [mailto:computer-go-boun...@computer-go.org] On Behalf > Of Gian-Carlo Pascutto > Sent: Wednesday, October 18, 2017 4:33 PM > To: computer-go@computer-go.org > Subject: Re: [Computer-go] AlphaGo Zero > > On 18/10/2017 19:50, cazen...@ai.univ-paris8.fr wrote: > > > > https://deepmind.com/blog/ > > > > http://www.nature.com/nature/index.html > > Select quotes that I find interesting from a brief skim: > > 1) Using a residual network was more accurate, achieved lower error, and > improved performance in AlphaGo by over 600 Elo. > > 2) Combining policy and value together into a single network slightly > reduced the move prediction accuracy, but reduced the value error and > boosted playing performance in AlphaGo by around another 600 Elo. > > These gains sound very high (much higher than previous experiments with > them reported here), but are likely due to the joint training. > > 3) The raw neural network, without using any lookahead, achieved an Elo > rating of 3,055. ... AlphaGo Zero achieved a rating of 5,185. > > The increase of 2000 Elo from tree search sounds very high, but this may > just mean the value network is simply very good - and perhaps relatively > better than the policy one. (They previously had problems there that SL > > RL for the policy network guiding the tree search - but I'm not sure > there's any relation) > > 4) History features Xt; Yt are necessary because Go is not fully > observable solely from the current stones, as repetitions are forbidden. > > This is a weird statement. Did they need 17 planes just to check for ko? > It seems more likely that history features are very helpful for the > internal understanding of the network as an optimization. That sucks though > - it's annoying for analysis and position setup. > > Lastly, the entire training procedure is actually not very complicated at > all, and it's hopeful the training is "faster" than previous approaches - > but many things look fast if you can throw 64 GPU workers at a problem. > > In this context, the graphs of the differing network architectures causing > huge strength discrepancies are both good and bad. Making a better pick can > cause you to get massively better results, take a bad pick and you won't > come close. > > -- > GCP > ___ > 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 > ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo Zero
So I am reading that residual networks are simply better than normal convolutional networks. There is a detailed write-up here: https://blog.waya.ai/deep-residual-learning-9610bb62c355 Summary: the residual network has a fixed connection that adds (with no scaling) the output of the previous level to the output of the current level. The point is that once some layer learns a concept, that concept is immediately available to all downstream layers, without need for learning how to propagate the value through a complicated network design. These connections also provide a fast pathway for tuning deeper layers. -Original Message- From: Computer-go [mailto:computer-go-boun...@computer-go.org] On Behalf Of Gian-Carlo Pascutto Sent: Wednesday, October 18, 2017 4:33 PM To: computer-go@computer-go.org Subject: Re: [Computer-go] AlphaGo Zero On 18/10/2017 19:50, cazen...@ai.univ-paris8.fr wrote: > > https://deepmind.com/blog/ > > http://www.nature.com/nature/index.html Select quotes that I find interesting from a brief skim: 1) Using a residual network was more accurate, achieved lower error, and improved performance in AlphaGo by over 600 Elo. 2) Combining policy and value together into a single network slightly reduced the move prediction accuracy, but reduced the value error and boosted playing performance in AlphaGo by around another 600 Elo. These gains sound very high (much higher than previous experiments with them reported here), but are likely due to the joint training. 3) The raw neural network, without using any lookahead, achieved an Elo rating of 3,055. ... AlphaGo Zero achieved a rating of 5,185. The increase of 2000 Elo from tree search sounds very high, but this may just mean the value network is simply very good - and perhaps relatively better than the policy one. (They previously had problems there that SL > RL for the policy network guiding the tree search - but I'm not sure there's any relation) 4) History features Xt; Yt are necessary because Go is not fully observable solely from the current stones, as repetitions are forbidden. This is a weird statement. Did they need 17 planes just to check for ko? It seems more likely that history features are very helpful for the internal understanding of the network as an optimization. That sucks though - it's annoying for analysis and position setup. Lastly, the entire training procedure is actually not very complicated at all, and it's hopeful the training is "faster" than previous approaches - but many things look fast if you can throw 64 GPU workers at a problem. In this context, the graphs of the differing network architectures causing huge strength discrepancies are both good and bad. Making a better pick can cause you to get massively better results, take a bad pick and you won't come close. -- GCP ___ 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
Re: [Computer-go] agz -- meditations
Sure, both hardware and software / algorithms are needed... but which gets you the bigger ROI? { Just a rhetorical question, I know it is not linear and not a simple question... but in general, I can see David Silver's (& Richard Lorentz / Demis Hassabis' counter) point }. May you live in sente, Cyris My AGA Go rank: http://agagd.usgo.org/player/13530/ Brian Cloutier replied: > Well, if you have both, why not use both :) > > On Oct 19th 2017 Richard Lorentz wrote: > >> An interesting juxtaposition. >> >> Silver said "algorithms matter much more than ... computing". >> >> Hassabis estimated they used US$25 million of hardware. >> ___ >> Computer-go mailing list > > ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo Zero
I would like to know how much handicap the Master version needs against the Zero version. It could be less than black without komi or more than 3 stones. Handicap differences cannot be deduced from regular Elo rating differences, because it varies depending on skill (a handicap stone is more than 100 regular Elo points points for higher dan players and less than 100 regular Elo points for kyu players). Dave de Vos >Origineel Bericht-- -- >Van : 3-hirn-ver...@gmx.de >Datum : 19/10/2017 20:53 >Aan : computer-go@computer-go.org >Onderwerp : Re: [Computer-go] AlphaGo Zero > >What shall I say? >Really impressive. >My congratulations to the DeepMind team! > >> https://deepmind.com/blog/ >> http://www.nature.com/nature/index.html > >* Would the same approach also work for integral komi values >(with the possibility of draws)? If so, what would the likely >correct komi for 19x19 Go be? > >* Or in another way: Looking at Go on NxN board: >For which values of N would the DeepMind be confident >to find the correct komi value? > > >* How often are there ko-fights in autoplay games of >AlphaGo Zero? > >Ingo. > >PS(a fitting song). The opening theme of >Djan-Go Unchained (with a march through a desert of stones): >https://www.youtube.com/watch?v=R1hqn8kKZ_M >___ >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
Re: [Computer-go] agz -- meditations
Well, if you have both, why not use both :) On Thu, Oct 19, 2017 at 11:51 AM Richard Lorentz wrote: > An interesting juxtaposition. > > Silver said "algorithms matter much more than ... computing". > > Hassabis estimated they used US$25 million of hardware. > ___ > 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
Re: [Computer-go] AlphaGo Zero
What shall I say? Really impressive. My congratulations to the DeepMind team! > https://deepmind.com/blog/ > http://www.nature.com/nature/index.html * Would the same approach also work for integral komi values (with the possibility of draws)? If so, what would the likely correct komi for 19x19 Go be? * Or in another way: Looking at Go on NxN board: For which values of N would the DeepMind be confident to find the correct komi value? * How often are there ko-fights in autoplay games of AlphaGo Zero? Ingo. PS(a fitting song). The opening theme of Djan-Go Unchained (with a march through a desert of stones): https://www.youtube.com/watch?v=R1hqn8kKZ_M ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
[Computer-go] agz -- meditations
An interesting juxtaposition. Silver said "algorithms matter much more than ... computing". Hassabis estimated they used US$25 million of hardware. ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo Zero
Yes, it seems really odd that they didn't add a plane of all ones. The "heads" have weights that depend on the location of the board, but all the other layers can't tell the difference between a lonely stone at (1,1) and one at (3,3). In my own experiments (trying to predict human moves) I found that 3 inputs worked well: signed liberties, age capped at 8, all ones. I think of the number of liberties as a key part of the game mechanics, so I don't think it detracts from the purity of the approach, and it's probably helpful for learning about life and death. Álvaro. On Thu, Oct 19, 2017 at 7:42 AM, Gian-Carlo Pascutto wrote: > On 18-10-17 19:50, cazen...@ai.univ-paris8.fr wrote: > > > > https://deepmind.com/blog/ > > > > http://www.nature.com/nature/index.html > > Another interesting tidbit: > > The inputs don't contain a reliable board edge. The "white to move" > plane contains it, but only when white is to move. > > So until AG Zero "black" learned that a go board is 19 x 19, the white > player had a serious advantage. > > I think I will use 18 input layers :-) > > -- > GCP > ___ > 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
Re: [Computer-go] AlphaGo Zero
The order of magnitude matches my parameter numbers. (My attempt to reproduce a simplified version of this is currently evolving at https://github.com/pasky/michi/tree/nnet but the code is a mess right now.) On Thu, Oct 19, 2017 at 07:23:31AM -0400, Álvaro Begué wrote: > This is a quick check of my understanding of the network architecture. > Let's count the number of parameters in the model: > * convolutional block: (17*9+1)*256 + 2*256 > [ 17 = number of input channels >9 = size of the 3x3 convolution window >1 = bias (I am not sure this is needed if you are going to do batch > normalization immediately after) > 256 = number of output channels >2 = mean and standard deviation of the output of the batch normalization > 256 = number of channels in the batch normalization ] > * residual block: (256*9+1)*256 + 2*256 + (256*9+1)*256 + 2*256 > * policy head: (256*1+1)*2 + 2*2 + (2*361+1)*362 > * value head: (256*1+1)*1 + 2*1 + (1*361+1)*256 + (256+1)*1 > > Summing it all up, I get 22,837,864 parameters for the 20-block network and > 46,461,544 parameters for the 40-block network. > > Does this seem correct? > > Álvaro. > > > > On Thu, Oct 19, 2017 at 6:17 AM, Petr Baudis wrote: > > > On Wed, Oct 18, 2017 at 04:29:47PM -0700, David Doshay wrote: > > > I saw my first AlphaGo Zero joke today: > > > > > > After a few more months of self-play the games might look like this: > > > > > > AlphaGo Zero Black - move 1 > > > AlphaGo Zero White - resigns > > > > ...which is exactly what my quick attempt to reproduce AlphaGo Zero > > yesterday converged to overnight. ;-) But I'm afraid it's because of > > a bug, not wisdom... > > > > Petr Baudis > > ___ > > 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 -- Petr Baudis, Rossum Run before you walk! Fly before you crawl! Keep moving forward! If we fail, I'd rather fail really hugely. -- Moist von Lipwig ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo Zero
On 18-10-17 19:50, cazen...@ai.univ-paris8.fr wrote: > > https://deepmind.com/blog/ > > http://www.nature.com/nature/index.html Another interesting tidbit: The inputs don't contain a reliable board edge. The "white to move" plane contains it, but only when white is to move. So until AG Zero "black" learned that a go board is 19 x 19, the white player had a serious advantage. I think I will use 18 input layers :-) -- GCP ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo Zero
This is a quick check of my understanding of the network architecture. Let's count the number of parameters in the model: * convolutional block: (17*9+1)*256 + 2*256 [ 17 = number of input channels 9 = size of the 3x3 convolution window 1 = bias (I am not sure this is needed if you are going to do batch normalization immediately after) 256 = number of output channels 2 = mean and standard deviation of the output of the batch normalization 256 = number of channels in the batch normalization ] * residual block: (256*9+1)*256 + 2*256 + (256*9+1)*256 + 2*256 * policy head: (256*1+1)*2 + 2*2 + (2*361+1)*362 * value head: (256*1+1)*1 + 2*1 + (1*361+1)*256 + (256+1)*1 Summing it all up, I get 22,837,864 parameters for the 20-block network and 46,461,544 parameters for the 40-block network. Does this seem correct? Álvaro. On Thu, Oct 19, 2017 at 6:17 AM, Petr Baudis wrote: > On Wed, Oct 18, 2017 at 04:29:47PM -0700, David Doshay wrote: > > I saw my first AlphaGo Zero joke today: > > > > After a few more months of self-play the games might look like this: > > > > AlphaGo Zero Black - move 1 > > AlphaGo Zero White - resigns > > ...which is exactly what my quick attempt to reproduce AlphaGo Zero > yesterday converged to overnight. ;-) But I'm afraid it's because of > a bug, not wisdom... > > Petr Baudis > ___ > 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
Re: [Computer-go] AlphaGo Zero
On Thu, Oct 19, 2017 at 11:04 AM, Hiroshi Yamashita wrote: > I have two questions. > > 2017 Jan, Master , defeat 60 pros in a row. > 2017 May, Master?, defeat Ke Jie 3-0. > > Master is Zero method with rollout. > Zero is Zero method without rollout. > > Did AlphaGo that played with Ke Jie use rollout? > Is Zero with rollout stronger than Zero without rollout? > Hi Hiroshi, I think these are good questions. You can ask them at https://www.reddit.com/r/MachineLearning/comments/76xjb5/ama_we_are_david_silver_and_julian_schrittwieser/ Aja > Thanks, > Hiroshi Yamashita > > - Original Message - From: > To: > Sent: Thursday, October 19, 2017 2:50 AM > Subject: [Computer-go] AlphaGo Zero > > > >> https://deepmind.com/blog/ >> >> http://www.nature.com/nature/index.html >> >> Impressive! >> > > ___ > 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
Re: [Computer-go] AlphaGo Zero
On Wed, Oct 18, 2017 at 04:29:47PM -0700, David Doshay wrote: > I saw my first AlphaGo Zero joke today: > > After a few more months of self-play the games might look like this: > > AlphaGo Zero Black - move 1 > AlphaGo Zero White - resigns ...which is exactly what my quick attempt to reproduce AlphaGo Zero yesterday converged to overnight. ;-) But I'm afraid it's because of a bug, not wisdom... Petr Baudis ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo Zero
I have two questions. 2017 Jan, Master , defeat 60 pros in a row. 2017 May, Master?, defeat Ke Jie 3-0. Master is Zero method with rollout. Zero is Zero method without rollout. Did AlphaGo that played with Ke Jie use rollout? Is Zero with rollout stronger than Zero without rollout? Thanks, Hiroshi Yamashita - Original Message - From: To: Sent: Thursday, October 19, 2017 2:50 AM Subject: [Computer-go] AlphaGo Zero https://deepmind.com/blog/ http://www.nature.com/nature/index.html Impressive! ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go