Chessbase Interview

[Javaness2]

I noticed a few days ago that there was an interview related to Leela Zero, the Chess variant.
https://en.chessbase.com/post/leela-che ... for-the-pc
The interview promises an upcoming interview with Gian-Carlo Pascutto to be published on the same site.

The way the interviewer tested the strength of the machine is interesting. It seems that according to some classic benches he was expecting it to get crushed by some of the weaker Chess engines. In fact, this did not happen, and they go on to explain why.

[EdLee]

Java, Thanks.

Nevertheless, it was believed that while a game such as Go, might fit this form of software engineering, it would never really work for a chess program. At least not at the highest level. After all, Go is less about calculating exact lines, and more about extensive pattern recognition, while chess is highly tactical and seeing just a single extra move ahead can make or break a program.

(emphasis added) Hmm...

[dfan]

If I had to pick one game as being more tactical and one as being more strategic, I would make the same choices as the article. Of course both games contain plenty of both elements.

[Calvin Clark]

With computer chess the focus has been on calculating. Forced sequences that can end the game are common. In Go, there are forced sequences that can make you feel sad, but reading to the very end is not possible until you are already close to the end.

But the amateur feeling of losing games based on tactical blunders may be similar in both. There are abstract characteristics of the game, and then there is how we experience it with our limited time and skill.

The author is shocked that Leela chess won so much while being bad at tactics. In go, we are puzzled that a program can apparently function at pro level without reading ladders. It is puzzling.

Lee Changho used to give the standard advice to study tesuji, life and death and things that make you read. Shortly after AlphaGo he started saying maybe we should get better at positional judgement.

[gowan]

As I recall, Alpha-zero, essentially the same methods of learning as AlphGO-zero, trounced the "calculating" programs like Stockfish. That would suggest that there is nothing in these comments suggesting how LeelaZero could not learn to play chess at a very high level. As for pattern recognition in chess, there is a lot of it, including such things as weak squares.

[Bill Spight]

About Alpha Zero (chess) trouncing Stockfish: Stockfish was hobbled. No opening book, not the latest version, no endgame database (IIRC), and maybe not an appropriately strong computer to run on. How come?

Yes, Alpha Zero was a major accomplishment and an inspiration. But I suspect that it had leveled off, so that another four hours, or even four days of self play would not have brought it up to the level of the top chess engines.

[dfan]

About Alpha Zero (chess) trouncing Stockfish: Stockfish was hobbled. No opening book, not the latest version, no endgame database (IIRC), and maybe not an appropriately strong computer to run on. How come?

Yes, Alpha Zero was a major accomplishment and an inspiration. But I suspect that it had leveled off, so that another four hours, or even four days of self play would not have brought it up to the level of the top chess engines.

My opinion is different. We have (some) game records, not just the results, and Alpha Zero played some amazing moves that no one has refuted. It's too bad that Stockfish wasn't playing at full strength (I blame incompetence rather than malice, but who knows), but that doesn't the diminish the quality of the moves of Alpha Zero that we can observe.

Luckily, Leela Chess Zero is being developed, which will provide a lot more data points than the few we've been provided by DeepMind, which will help settle the argument one way or another.