Rise of the mighty NN

[Dann Corbit]

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.

[Dann Corbit]

SF has a few tricks up his fishy little sleeve, as he pulled out a very impressive draw

[Werewolf]

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.

Except in the endgame. Then it's the other way around.

I'd have expected NNs to be great in the endgame because of all the rules and principles but it doesn't seem to work.

[Dann Corbit]

Depth is the Achille's heel of NN programs.
In the endgame low branch factor programs like SF get absurd depths, while NN programs struggle mightily.

Look at the TB hits figures.

I like to put it this way:
In the end game, programs like SF spend most of their time talking to God (EGTB are flawless) while the NN engines fumble forward with merely superhuman strength.

[Werewolf]

Depth is the Achille's heel of NN programs.
In the endgame low branch factor programs like SF get absurd depths, while NN programs struggle mightily.

Look at the TB hits figures.

I like to put it this way:
In the end game, programs like SF spend most of their time talking to God (EGTB are flawless) while the NN engines fumble forward with merely superhuman strength.

Yes that's a poignant way of putting it.

However, would the AB still be better in the endgame if TBs were disabled? Assuming they would, what is the issue now?

[Dann Corbit]

the issue now is depth of search

[Werewolf]

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?

[carldaman]

Good question, Werewolf, I've often wondered about that myself. Maybe specific training on endings like what I assume dkappe is doing with Ender might be a solution.

[Nay Lin Tun]

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.

[Werewolf]

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.

Fair enough, but this was just a specific example.