TCEC season 13, 2 NN engines will be participating, Leela and Deus X

[jd1]

Well, there is extreme high possibility that Deus X authour is Deep Junior Authour, Shay Bushinsky.
https://www.aaai.org/ojs/index.php/aima ... /view/2255

So you will be seeing updated version(work of art) of Junior!!

No - you're wrong, but my lips are sealed.

What about Tencent company?
https://www.chess.com/forum/view/genera ... -season-13

http://www.chessdom.com/deus-x-the-nn-c ... rt-silver/

Meanwhile, Scorpio with neural networks is ignored ...

It uses supervised learning (just like the deuce) with different NN architecture, search, backend etc...

Hurts to be ignored after investing so much time in it ... is it because I am black (obviously I am joking here)

Keep up the good work, Daniel! I find Scorpio NN very interesting indeed, and judging from the other comments on here I'm far from alone.

[Graham Banks]

Meanwhile, Scorpio with neural networks is ignored ...

It uses supervised learning (just like the deuce) with different NN architecture, search, backend etc...

Hurts to be ignored after investing so much time in it ... is it because I am black (obviously I am joking here)

Hi Daniel,

I get the impression that both Scorpio and DisasterArea missed out this time due to having issues in TCEC last season.

[Damir]

It is no excuse Graham. The organizers of TCEC could have approached Daniel and asked him if he had a bugfix version of his engine. It would have than been up to him whether he wanted to participate or not...

[Albert Silver]

I think that there is a difference because I guess that Brainfish derivatives are similiar in the choice of moves when LeelaChess derivatives with a new net are not similiar in the choice of moves but I am not sure about it and it may be interesting to see the difference in move choice between LC0 and Deus and compare.

What the hell are you talking about???
Do you know how many lines of code of LC0 Albert changed? My guess is less than what a typical cloner of SF does.

Actually, exactly zero.

He has identical net,

No, the neural net is completely original, using training techniques, material, and other tweaks. Some ideas, which could still follow the 'zero' approach, I shared many times with Leela Devs, discussing to no avail, others not.

identical number of inputs, levels, filters, using exactly the same backend, same search.
What he did is identical to taking SF, changing PST numbers and evaluation bonuses (like what Tsvetkov was doing) and claiming the new engine that is btw. probably hundreds of Elo weaker than original SF. You really think that changing only some numbers randomly in SF eval wouldn't make it play totally different moves?

Changing random numbers, as you suggest, would mean using some NN of Leela, instead of toiling on this for months as I have, building it from scratch, with numerous stalls and restarts along the way. Building a NN isn't that hard, but building a good one, much less a really good one, is very hard.

And after all there is question of legality, since TCEC is public event and that "Deus X" competes on it, if he doesn't publish the source code he is in direct violation of not only GPL licence but also NVIDIA licence.

Source code? Of lc0? I can give you a link if you want.

If you really want to see innovative NN engine look at Scorpio MCTS+NN of Daniel. Totally different net, backend, search, everything.

I thought he was already playing. If not, I agree it would be a terrible shame for him not to.

[frankp]

Albert, what is your source of games to train the Leela network with?
Just interested, since Leela is now trained against >20M games and this has taken months of around 200 contributors.

[Laskos]

I think this move may backfire and will hurt TCEC more than it will do good. People are not stupid and want diversity, not 32 clones of SF. Deus X will use lc0 with a custom net. Shouldn't be allowed imho.

I agree, it seems rather difficult to me to justify allowing Deus X while not letting developers of other strong engines enter clones with changed parameters under a new name. Let's see how TCEC spins this once the complaints to this move start coming in...

For Season 13, it seems the decision is made, but for the season after this they'll have to come up with a new set of rules and follow it consistently.

I don't agree. One uses supervised learning, another reinforcement learning. They will play completely different chess. Also, as the 2 groups (main and test) seem to stall at significantly below A0 level (even if put on the same hardware), supervised learning is worth give a try. I hope this Deus X wins the TCEC, and all these discussions will be put to rest.

It might be worth a try for fun, we also did it in the initial phase of lc0 development, see https://github.com/glinscott/leela-ches ... -361063249. We didn't have the fast engine with cudnn then and limited ourselves to 64x6. It was done to make sure the actual network architecture was sane (not just weights) .

I think it's very unlikely that this approach will beat A0 ever. It's based on human games, with its human flaws and limitations. It'll obtain generalized knowledge across many humans which is of course better, but it will still not see beyond human capability (given just a nn forward inference, with MCTS combined it will somewhat).

A nn trained through Selfplay from scratch, however won't be bound by humans. And looking at deepmind results for Go, it far surpassed the supervised approach. We're (lc0) just struggling/learning to find the optimal parameters for training and avoiding overfitting, as it's much more challenging to get right (and actually write the entire engine etc). But do note that our 20b net is still in it's infancy, it'll become interesting as it drops learning-rate and starts fine-tuning its performance.

Thanks, I didn't know you tried supervised learning as sanity check. How was it performing at tactics? One of the things which bother me is the dismal tactical abilities of all the attempts so far. And the regression in these with the new nets, especially in the main branch. Perhaps sometime in the future one has to abandon the pure UCT MCTS search for some hybrids with AB in the manner Daniel exemplified.

The initial AlphaGo was trained using some server human games, maybe as weak as KGS games (IIRC). But it played at about 1000 Elo points above that level against Lee Sedol (well, on a monster hardware). Chess has compressed Elo differences compared to Go, so supervised learning might give 500 Elo points better results than say training games of averaging 2500 level humans, on a strong hardware. 3000 FIDE Elo points is perhaps achievable with supervised learning (on a strong hardware). Current main- and test- server nets on a strong hardware are considerably stronger than that, say 3200 FIDE Elo level. So, yes, it does seem plausible that Deus X will be weaker in TCEC than Lc0 with a reasonable net of the developers using reinforcement learning. I think 20b nets from testserver scale better to LTC and hardware than 15b mainserver nets, so in TCEC conditions those 20b nets should be preferred, maybe in 10160-10190 range of IDs.

Another issue: it was going so well with 6x64 nets on testserver, 15 times faster games, quick saturation at high Elo value (surely above 3100 CCRL 40/4'). They only scaled badly with TC, but that's understandable. Wasn't it simpler to proceed 6b -> 10b -> 15b -> 20b? I guess even with 10b nets we would have had stronger nets than the current ones, and it was real fast, a matter of days. Then to slower 15b and to slowish and seemingly tricky 20b. I don't even know how one would tune CPUCT, FPU and such, as in my tries their optima strength-wise depend heavily on time control (and hardware) used.

[jorose]

Albert, what is your source of games to train the Leela network with?
Just interested, since Leela is now trained against >20M games and this has taken months of around 200 contributors.

I would imagine one of the primary reasons he is using "Human" games (I thought I read somewhere he is using correspondence games? If so, hardly human imo) is precisely because it uses several orders of magnitude less resources. Most of the resources required from those contributors were to create self play games, sometimes between absolute garbage quality networks. Each network is trained on a subset of those games, not all of them. If you do not have insane resources, like a distributed project or large company can have, then going the zero approach is not really feasible at the moment.

On the other hand I am not convinced you need too many games in order to train a network capable of producing much higher quality games than your dataset.

[frankp]

Albert, what is your source of games to train the Leela network with?
Just interested, since Leela is now trained against >20M games and this has taken months of around 200 contributors.

I would imagine one of the primary reasons he is using "Human" games (I thought I read somewhere he is using correspondence games? If so, hardly human imo) is precisely because it uses several orders of magnitude less resources. Most of the resources required from those contributors were to create self play games, sometimes between absolute garbage quality networks. Each network is trained on a subset of those games, not all of them. If you do not have insane resources, like a distributed project or large company can have, then going the zero approach is not really feasible at the moment.

On the other hand I am not convinced you need too many games in order to train a network capable of producing much higher quality games than your dataset.

Thanks.
I found some information on the leela forum site.
Seems an identical leela (LC0) net, but trained differently or on different data - I think.
Rather like a clone of, say, SF but with different pst.
Interesting to see how it performs, particularly if the (identical) leela net is trained on (a relatively) small set of only human games.

(Raises interesting questions about what the GPL licence applies to, should Albert decide to sell it as the interview linked on leela forum suggests may be a possibility. But that is a different subject.).

[Albert Silver]

Albert, what is your source of games to train the Leela network with?
Just interested, since Leela is now trained against >20M games and this has taken months of around 200 contributors.

I would imagine one of the primary reasons he is using "Human" games (I thought I read somewhere he is using correspondence games? If so, hardly human imo) is precisely because it uses several orders of magnitude less resources. Most of the resources required from those contributors were to create self play games, sometimes between absolute garbage quality networks. Each network is trained on a subset of those games, not all of them. If you do not have insane resources, like a distributed project or large company can have, then going the zero approach is not really feasible at the moment.

On the other hand I am not convinced you need too many games in order to train a network capable of producing much higher quality games than your dataset.

Thanks.
I found some information on the leela forum site.

If you mean Chris Whittington's fountain of wisdom, I would question the use of the word 'information'.

[jorose]

[...]

[...]

[...]

[...]

Thanks, I didn't know you tried supervised learning as sanity check. How was it performing at tactics? One of the things which bother me is the dismal tactical abilities of all the attempts so far. And the regression in these with the new nets, especially in the main branch. Perhaps sometime in the future one has to abandon the pure UCT MCTS search for some hybrids with AB in the manner Daniel exemplified.

The initial AlphaGo was trained using some server human games, maybe as weak as KGS games (IIRC). But it played at about 1000 Elo points above that level against Lee Sedol (well, on a monster hardware). Chess has compressed Elo differences compared to Go, so supervised learning might give 500 Elo points better results than say training games of averaging 2500 level humans, on a strong hardware. 3000 FIDE Elo points is perhaps achievable with supervised learning (on a strong hardware). Current main- and test- server nets on a strong hardware are considerably stronger than that, say 3200 FIDE Elo level. So, yes, it does seem plausible that Deus X will be weaker in TCEC than Lc0 with a reasonable net of the developers using reinforcement learning. I think 20b nets from testserver scale better to LTC and hardware than 15b mainserver nets, so in TCEC conditions those 20b nets should be preferred, maybe in 10160-10190 range of IDs.

Another issue: it was going so well with 6x64 nets on testserver, 15 times faster games, quick saturation at high Elo value (surely above 3100 CCRL 40/4'). They only scaled badly with TC, but that's understandable. Wasn't it simpler to proceed 6b -> 10b -> 15b -> 20b? I guess even with 10b nets we would have had stronger nets than the current ones, and it was real fast, a matter of days. Then to slower 15b and to slowish and seemingly tricky 20b. I don't even know how one would tune CPUCT, FPU and such, as in my tries their optima strength-wise depend heavily on time control (and hardware) used.

Honestly I have a strong suspicion the tactical weakness of Leela is due to the google algorithm and not due to a weakness of the neural net. A smaller neural net could help a lot in that respect, simply allowing more nodes to be searched at the cost of precision. It is also not clear to me why UCT search should be better than a modified AB search, such as Matthew Lai's probability based search, once the network has been trained, that's actually something I want to try with Leela when I have time, but with my thesis deadline coming up and me enjoying working on Winter I don't think that will happen in the near future.

I believe the original AlphaGo was only bootstrapped with human games. It was not trained in a supervised fashion after that iirc. The games are very different though, it is quite possible that in chess supervised training will work very well, even if it doesn't in Go.