Leela starts to beat medium class engines

The result of a 20 game match Match Yace Paderborn (Elo ~ 2500) versus Leela was 18.5 - 1.5. It was 5 minutes in 40 moves, both had one processor. GUI was Cute Chess (Im am not able to get Leela running on Scid.)

Here is Leelas only win:



Like Alpha Zero, Leela likes intuitive sacs.

Support Leela here:

github.com/glinscott/leela-chess/wiki/Getting-Started

BTW, if Leela had 50 000 supporters, then Leela would become the strongest chess engine in existence in less than three days. Currently it has 500+, it will take a few years with that number...

here are some hints about running Leela as engine:

lichess.org/forum/game-analysis/leelazero#8

Ran about 300 games yesterday on my laptop. The fan was going crazy. I think you need serious GPUs to do more. Only 36 million games to go until it's close to Alpha Zero... And that's assuming it scales, the ELO rating seems to have flatlined in the last few days.

@jperkinz

I was asking myself same question about it flatlining. When comparing it with the Alpha Zero progress given in the Google paper, Alpha Zero also had a small period of flatlining. I dont know if or how that relates to the graphic on the Leela homepage.

5.g3 looks counterintuitive.

But looks like at depth 18, at least, Stockfish thinks it's OK. Very curious.

@tpsinnem

Leela is still in an early state of her knowledge. Weird errors are common in every game. This is the first game from this match:

@nh78 I'm not even qualified to say it's an error. As I said, Stockfish seems to also be OK with g3, at least at that relatively shallow depth.

Someone at your level might have a clearer idea, though. I'd actually be curious to hear it.

yes, it is playable. White has no advantage but black also hasnt. Can be classified as opening trap because when black takes on c2 white has good chances to play for an advantage (10 Kd1 +=). But white also doesnt risk anything except that it can be a quick perpetual:

I'm not sure 28. Rg8+ was an error. I think Leela knew she was lost after any normal move, but winning after Rg8+ unless black plays Bxg8. She must have thought there was a non-zero probability black would play Kd7 or Ke7.

Looking with Stockfish at what follows from 10.Kd1, it seems you can end up with two pawns for the exchange, but it's amazing that in seemingly none of these lines do you get to eventually just capture the knight on a1 without giving back some piece of your own.

Something always gets in the way.