LCZero Accomplishments and Goals Thus Far

[jp]

using the metric elo/byte ?

Indeed, but at source-code level. So Elo/character.

Not a metric that the average buyer of a Chess engine would give a hoot about, of course.

I am confident we can configure Stockfish using cutting-edge industry techniques to get the desired Min-Max match result for any metric the average buyer gives a hoot about.

[hgm]

Not sure what that means, and why anyone would give a hoot about it...

It doesn't seem related to the discussed topics at all.

[mar]

I don't care how Leela performs on a toaster, on my hw it seems to perform at ~2900 CCRL (ID 227) even at a fast 40 moves in 1 minute, no way it's 2100
Fully utilizing available hardware is hardly "unfair".

She will smoke 3k engines in next TCEC.

[Laskos]

I don't care how Leela performs on a toaster, on my hw it seems to perform at ~2900 CCRL (ID 227) even at a fast 40 moves in 1 minute, no way it's 2100
Fully utilizing available hardware is hardly "unfair".

She will smoke 3k engines in next TCEC.

Try 40 moves in 1 hour, it will be 3100 or so.

[lucasart]

I don't care how Leela performs on a toaster, on my hw it seems to perform at ~2900 CCRL (ID 227) even at a fast 40 moves in 1 minute, no way it's 2100
Fully utilizing available hardware is hardly "unfair".

She will smoke 3k engines in next TCEC.

Don't underestimate the toaster. With the internet of things, we might see toasters able to connect to large servers or bitcoin mining farms with enormous computing power…

Toasters could be the future of chess

[hgm]

As one robot said to the other: "I will explain this in such simple terms that even a washing machine can understand it!".

[FWCC]

We must also remember we will compare LCZero with the latest Stockfish Dev. Version not some watered down old version like DeepMind. We will be more in the know. In order to use mainstream hardware with success we must push Leela beyond A0.In time will she pass A0?

[hgm]

Oh, if the goal is just to make a strong engine, I am sure you can do much better than AlphaZero, algorithm wise. By not presenting the NN just with a raw position, but already pre-processing the position in a way known to be relevant for Chess, such as number of attackers / protectors of each square, the lowest attacker/protector of each square, pinned pieces and discovered attacks, etc. I am pretty sure it must recognize such elementary features to be any good, and then it would not have to learn recognizing them (in a way that even after it has learned it will be quite calculationally intensive).

Of course in practice you will also have to bridge the hardware gap between GPUs and TPUs.

[Milos]

Oh, if the goal is just to make a strong engine, I am sure you can do much better than AlphaZero, algorithm wise. By not presenting the NN just with a raw position, but already pre-processing the position in a way known to be relevant for Chess, such as number of attackers / protectors of each square, the lowest attacker/protector of each square, pinned pieces and discovered attacks, etc. I am pretty sure it must recognize such elementary features to be any good, and then it would not have to learn recognizing them (in a way that even after it has learned it will be quite calculationally intensive).

Lol, you are sure. When you are so sure, why don't you contribute with some code to LC0 git. Doesn't have to be main inefficient OpenCL code of Gian-Carlo, you can contribute to TF brench of it. Just implement your net in keras and add it to the repository so we can all test how strong it becomes. I'm pretty convinced ppl would test whatever you post no matter how much "out of spirit" of original A0 it was.

But somehow, I have a feeling you won't because you have no clue how. You know just to go around and splash your "wisdom".

[hgm]

I missed how much you are offering to pay me for that?

Actually I did play against a Shogi program that played by an NN + preprocessing along the lines I outlined above, and without any search it was able to beat my alpha-beta searcher at ~2min/move. (The NN needed ~10 sec/move, as it was playing on the CPU of a laptop.)

No reason to suppose the same couldn't work in Chess. But of course you are as ignorant about progress in Shogi as about almost anything else...