Intuitive Chess Match: Rebel Century CCRL Rated 2542 vs Lc0 Neural Network Play Only (0 Calculation)

Posts : 880 Join date : 2020-11-25 Location : USA

Intuitive Chess Match: Rebel Century CCRL Rated 2542 vs Lc0 Neural Network Play Only (0 Calculation)

Hardware: Threadripper 2950x, RTX 2080TI, 64 GBs Ram.

Book Perfect Book 2021 to 6 moves.
Rebel Century Default settings.
Lc0 setting is set to 1 node only to produce a move at each turn.
Lc0 is playing with zero calculation, and play is based on just the Neural networks
understanding of the position.
Time controls for Rebel Century. 2m+1s
Lc0 is playing instantly.

Poor Lc0

Posts : 880 Join date : 2020-11-25 Location : USA

Admin wrote:: Poor Lc0

Lc0 is doing great. Lc0 is playing a GM level chess engine, and is playing at a IM level with zero calculation.

I find that amazing.

Do not try this with any other NN like Stockfish. It will get CRUSHED!

Isn't Lc0 not doing a full depth=1, it certainly looks that way.

Posts : 880 Join date : 2020-11-25 Location : USA

Admin wrote:: Isn't Lc0 not doing a full depth=1, it certainly looks that way.

No it is set to 1 node. And just to produce the first move it sees.

So it is not even looking at all possible first moves. Just what it thinks would be at the top of the list.

Look at the node counter, and the PV. And ponder hits are 0%.

So no move are looked at for the other side. To see if what is played could just lose by tactics.

mwyoung wrote:

Admin wrote:: Isn't Lc0 not doing a full depth=1, it certainly looks that way.

No it is set to 1 node. And just to produce the first move it sees.

So it is not even looking at all possible first moves. Just what it thinks would be at the top of the list.

Look at the node counter, and the PV. And ponder hits are 0%.

So no move are looked at for the other side. To see if what is played could just lose by tactics.

Really, that sounds as a full depth=1 search.

But what are the engine settings you use?

Posts : 880 Join date : 2020-11-25 Location : USA

Admin wrote:

mwyoung wrote:

Admin wrote:: Isn't Lc0 not doing a full depth=1, it certainly looks that way.

No it is set to 1 node. And just to produce the first move it sees.

So it is not even looking at all possible first moves. Just what it thinks would be at the top of the list.

Look at the node counter, and the PV. And ponder hits are 0%.

So no move are looked at for the other side. To see if what is played could just lose by tactics.

Really, that sounds as a full depth=1 search.

But what are the engine settings you use?

I am not sure what the difference is.... You are lonely looking at your side, but not counter moves.

And in the banksia GUI in the time controls. You have the option to set how many nodes for the UCI chess engine to use instead of time.

So I use time controls MULTI option and set Rebels time to 2m+1s and Lc0 time to 1 node.

This is the weakest possible setting for Lc0.

Arena, nodes=1

Code:: FEN: rnbqkbnr/pppppppp/8/8/3P4/8/PPP1PPPP/RNBQKBNR b KQkq d3 0 1

Lc0:
  1/1 00:00 1 0 +0,04 e2-e4
  1/1 00:00 1 0 +0,04 c2-c4
  1/1 00:00 1 0 +0,04 Ng1-f3
  1/1 00:00 1 0 +0,04 d2-d4

Arena, depth=1

Code:: FEN: rnbqkbnr/pppppppp/8/8/3P4/8/PPP1PPPP/RNBQKBNR b KQkq d3 0 1

Lc0:
  1/1 00:00 1 0 +0,04 e2-e4
  1/1 00:00 1 0 +0,04 c2-c4
  1/1 00:00 1 0 +0,04 Ng1-f3
  1/1 00:00 1 0 +0,04 d2-d4

Posts : 880 Join date : 2020-11-25 Location : USA

Admin wrote:

Arena, nodes=1

Code:: FEN: rnbqkbnr/pppppppp/8/8/3P4/8/PPP1PPPP/RNBQKBNR b KQkq d3 0 1

Lc0:
  1/1 00:00 1 0 +0,04 e2-e4
  1/1 00:00 1 0 +0,04 c2-c4
  1/1 00:00 1 0 +0,04 Ng1-f3
  1/1 00:00 1 0 +0,04 d2-d4

Arena, depth=1

Code:: FEN: rnbqkbnr/pppppppp/8/8/3P4/8/PPP1PPPP/RNBQKBNR b KQkq d3 0 1

Lc0:
  1/1 00:00 1 0 +0,04 e2-e4
  1/1 00:00 1 0 +0,04 c2-c4
  1/1 00:00 1 0 +0,04 Ng1-f3
  1/1 00:00 1 0 +0,04 d2-d4

Correct, but nodes are not depth. Now set to 2 nodes. You will get 1 white move , and one black move. Now set to 5 nodes, this does not get you depth 5. It will get you depth 2, 3, 4 or 5. Depending the on the position, and how Lc0 thinks best to use the 5 nodes.

Lc0 set at depth 5 can use 100's of nodes in the calculation to reach depth 5.

Posts : 880 Join date : 2020-11-25 Location : USA

Code:: Result:
------------------------------------------------------------------------------------------------------------------
  # name games wins draws losses score elo + -
  1. Rebel Century CCRL Rating 2542 400 276 65 59 308.5 107 17 17
  2. Lc0 v0.29.0 Neural Network Only Play (0 calculation) 400 59 65 276 91.5 -107 17 17

Cross table:
------------------------------------------------------------------------------------------------------------------
  # name score games 1 2
  1. Rebel Century CCRL Rating 2542 308.5 400 x 308.5
  2. Lc0 v0.29.0 Neural Network Only Play (0 calculation) 91.5 400 91.5 x

Tech:
------------------------------------------------------------------------------------------------------------------

Tech (average nodes, depths, time/m per move, others per game), counted for computing moves only, ignored moves with zero nodes:
  # name nodes/m NPS depth/m time/m moves time
  1. Rebel Century CCRL Rating 2542 5676K 2481010 10.6 2.3 58.1 132.9
  2. Lc0 v0.29.0 Neural Network Only Play (0 calculation) 0K 29 1.0 0.0 59.6 2.0
   all --- 2735K 2443893 5.7 1.1 58.9 67.5

Posts : 116 Join date : 2022-07-19

1 node on Leela is just policy IIRC, real depth 1 and depth 2 on leela take a little longer

Posts : 1254 Join date : 2020-11-17 Location : France

texium wrote:: 1 node on Leela is just policy IIRC, real depth 1 and depth 2 on leela take a little longer

I think that's correct. Policy is the move selector not the evaluator, anymore than AB search killer/hash are evaluators.

Depth 1 for AB looks at all moves plus Qsearch captures (eg more nodes than just the ply one width). MCTS looks at all moves at ply one(?) but each is extended. Probably MCTS d1 looks at more nodes than AB d1, but what can you do? Play nodes=100 matches? Nodes 50?

Posts : 116 Join date : 2022-07-19

Yea it's interesting. A lot of Leela testing does 50, 100, 500 nodes because it's perfect for LTC type testing. MCTS takes less nodes overall and from it's readme file, Dragon with MCTS outputs an estimation of depth

Posts : 116 Join date : 2022-07-19

lc0 doesn't use depth actually, its like fake depth

» How strong can the smartest chess engine in the world play intuitive chess?
» The dumbest neural network there is
» The Year Computer Chess Overtook The Very Best Human Chess Play.
» Live Chess 960 Match Dragon by Komodo Chess vs Stockfish 291120 (TC=90m+30s)(32 Threads)
» Play online chess with real chess board