Rise of the mighty NN

[Ovyron]

Why not make a Leela Chess One? Lc1 would use the normal approach for most of the game, then switch to an A/B approach on the problematic endgame. Best of both worlds.

Laskos has shown this could be the strongest thing and there's at least 2 different tools that do this, yet we don't get to see it used on a tournament, for some reason.

[Milos]

the issue now is depth of search

OK, but consider the ending K+B+N v K which some have said causes NN trouble (I haven't tested this btw)

My Mephisto Vancouver 68000 barely managed 1000 n/s and performs this endgame very well.
I myself barely manage 1 n/s and look hardly 3-5 plies ahead but I can win this in my sleep.

I would have thought a pattern recognising NN would also win with no trouble.

Is the issue that in the search (rollouts?) the NN sees that a high % of its internal games win which masks there may be a drawing line?

I am pretty sure current Leela can easily mate KBN vs K endgame without tablebases.( Current SV Leela is a super feeded training data from combination of T40( 60 millions games) +T60 + further training).

Afaik, Leela could do KNB vs K checkmate around 5~10 million training since T40.

At what nps? I bet at 10nps it wouldn't be able.

[Milos]

Right now, AllieStein has taken Stockfish out behind the woodshed and given the poor fishy the business. He put the fish on a stick, smoked it with some nice hickory wood, and started chewing on his ribs.

We are more than a few games into the tournament. If mighty SF wants to finish on the podium, he had better swim upstream a little harder.

The NN engines bear their mighty arm. The AB engines signal loud alarm. Maybe they will come to harm. But the NN guys are coming, in a swarm.

Somehow I feel you are gonna feel stupid for your comments at the end of current TCEC stage.
What we currently have is 3 engines with identical search and different eval. Now imagine 3 copies of SF all with identical search and only changed eval. How many NN engines would be in top 3 then?

[Dann Corbit]

You are correct in your assertion that multiple NN engines gives them enhanced winning chances.

If one million stockfish engines were entered in a contest and there was one NN engine of about equal ability, the NN engine would lose. Especially in a short contest, an engine may play 100 Elo above or below its true strength.

I don't know if "stove meat" has had a radical improvement or if it has just been fortunate.

One still must view the rise of NN engines with a bit of awe no matter what. The successful AB engine has been going strong for decades. The first successful NN engine was spawned from Matthew Lai's project (as soon as they threw a bazillion horsepower at the idea in the Alpha Zero project) and so the success was startling.

[Milos]

One still must view the rise of NN engines with a bit of awe no matter what. The successful AB engine has been going strong for decades. The first successful NN engine was spawned from Matthew Lai's project (as soon as they threw a bazillion horsepower at the idea in the Alpha Zero project) and so the success was startling.

Metthew Lai's project impact on Alpha Zero is in the same ballpark as HGMs Micro-Max on SF.
A0 search and training all come from Alpha Go Zero, and for NN model used in A0 and Metthew's engine only common point is that they are CNN. And that's about it.
The real reason behind the success for NN engines is fast matrix multiplication enabled on modern GPUs by cutting-edge litographic processes (14nm and below) and on dedicated AI hardware (such as TPUs). 20 years ago we didn't even have computational resources to handle NN with more than a single hidden layer. In the same time we had Cray and IBM's Deep Blue.

[brianr]

The overall impact of Giraffe should not be overlooked.

Firstly, it demonstrated that NN engines could work reasonably well, even without powerful GPUs.
This in turn was exciting enough to rekindle interest in computer chess, which had become a rather torpid field by 2015. This resurgent interest in CC exploded with the appearance of Alpha Zero in Dec 2017.
And, Matthew was a member of the DeepMind AZ team.

So, the impact was quite a bit more than MicroMax relative to Stockfish, although perhaps a bit less than the impact of Glaurung leading to SF.

[Dann Corbit]

You do realize that Matthew Lai was one of the key programmers on the Alpha Zero project.

His CPU engine played fairly well, despite an incredible burden of floating point more taxing than a CPU wants to bear. I suggested switching to GPUs long before the Alpha Zero project (but Matthew was not too keen on it at the time).

It is his fundamental algorithm and vision that drove the entire project. His method is not like the previous TD leaf attempts at all.

And while HGM's tiny chess program is clever, I do not see any enormous impact on the Stockfish project. So I would have to call that an exaggeration.

[Milos]

Oh, for god sake, how is it that you know so little about history of AlphaGoZero or implementation for the matter???
AlphaZero is merely a tweak of AlphaGoZero. And AlphaGoZero is derived from AlphaGo the difference being not in search or NN implementation but in training. AlphaGo was out much before Metthew joing DeepMind (I am not even sure he joined DeepMind before AlphaGoZero was in beta stage, i.e. AlphaGo Master time). And no he was nothing like a key programmer, just one screw in the machinery out of more than 20 ppl working on the project in its later stage.

[brianr]

There is no question that significant work was being done with NN engines for Go before Giraffe.
Nonetheless, the crossover from Go back to Chess using both NNs and the "zero" approach re-energized interest in computer chess. While Alpha Zero for chess would have happened anyway, I for one am grateful to Matthew for Giraffe and sharing his work on what was a relatively radical idea within the domain of chess.

Moreover, it is interesting to note the crossover between Go and Chess: Gian-Carlo Pascutto (author of Faile and Sjeng) working on Leela Go, and then again back to chess with Stoofvlees (NN); and Gary Linscott working on Fishtest and then the original Leela Chess Zero leveraging GCP's Leela Go; Daniel Shawul with Scorpio (and its own tablebases) and now Scorpio-NN.

[Dann Corbit]

Giraffe was released in July 2015.
First appearance of AlphaGo October 2015.
The fact that they hired him shows they thought his ideas were good.