HIARCS 14 vs LCZero

[Werewolf]

I'm playing 20 game matches between Leela Chess Zero and HIARCS.

As LCZero progresses, I'll try to beef HIARCS up by adding more cores etc.

Conditions:

HIARCS 14 WCSC
One core, Intel Broadwell 4.2 GHz
Built in book enabled

LCZero ID 202
Nvidia 1060 Graphics Card
No Book

Noomen Short Lines Testsuite, first 10 positions, played twice with reversed colours

Time Control 5 mins per game + 5 sec increment.


The match ended as a win for HIARCS, final score:

12.5 - 7.5

HIARCS 14 scored
9 wins
7 draws
4 losses

A month ago this match would have been unthinkable, as LCZero was still struggling against Dedicated machines from the 1990s!

I'll run the match again next weekend under the same conditions.

Anyone want to guess the score???

[Martin Hertz]

I'll run the match again next weekend under the same conditions.

Anyone want to guess the score???

Not easy to guess, maybe HIARCS will score 11-9 next week.

[Werewolf]

I'll run the match again next weekend under the same conditions.

Anyone want to guess the score???

Not easy to guess, maybe HIARCS will score 11-9 next week.

Yes I think that would be my guess as well. Let's see.

[Jhoravi]

Please limit the opening length to 2 moves or 4 plies. It's more interesting to see the engines doing the most part of it by themselves.

[Werewolf]

Please limit the opening length to 2 moves or 4 plies. It's more interesting to see the engines doing the most part of it by themselves.

I wanted this match to be a bit different. I think I remember long ago the HIARCS team made the point that the opening book is part of the program, in the sense that they've bothered to make it and include it in the final package.

I wanted this to be about two entities, one well prepared, crafted and polished by humans, the other self-learning with no outside help.

[Werewolf]

Same conditions as before:

LCZero 321 vs HIARCS 14

5 wins
6 draws
9 losses

8 - 12

HIARCS wins again. But for how much longer?

[Werewolf]

I installed Nvidia CUDA on my machine, as well as the necessary dlls, and got LCZero running on that.

What a difference. Node count is about 3x higher on average (so 3000 nps woop woop) and I was using version 349. All other settings the same as before and no hardware change. HIARCS was calculating at about 800,000 nps.

LCZero ID 349 CUDA - HIARCS 14
10 wins
6 draws
4 losses

to LCZero. A very convincing win.