lichess.org
Donate

Rating of AlphaZero?

Anyone know what the approximate rating of AlphaZero is? 3600+?
I heard it played StockFish 100 games, didn't lose a single game, won 28, drew 72.
It has broken the rating system as it has yet to be beaten.
Those knowledgeable on the subject, after review of the games estimate at the high end a rating of 2700+. This rating is very generous, based on the severe handicaps SF had to play with.
LOL. The guy in the video is a dufass. Hasn't a clue that SF played under severe handicaps and was nowhere near its full strength that he based his calculations on.
4100+. What a hoot. The video guy clearly has no understanding of ratings and the progression inherent in every 100 points. If A0 won every game, he'd calculate a rating of 10,000!
Point is, if a program won every game, may as well say it is rated 1,000,000. Why not?
It seems a rating of 4000 is not unreasonable, given how many times AZ beat SF with so few losses. Amazing.
Lets assume Stockfish was 3400 (and apparently it was weaker in the touranment?)

Then 3500-3600 would be a reasonable guess for alpha zero.

There are charts out there which can roughly estimate these things.

You can even see the performance of how stockfish does with weaker engines and check the rating difference.

Not sure that makes sense but ya.
Because AlphaZero can easily improve just by playing any opponent or itself, it does not have a fixed rating. I believe an estimated rating after its first 4 hours of training were ranging from 3600 to 4000, well exceeding the strength of human masters.
In the AlphaZero paper they have some charts showing an approximate ELO rating between the engines. According to that, his ELO is like 100 points higher than SF's. Remember that in chess most games are draws, and this mitigates the difference. That is not the case with shogi and Go.

Also, ELO ratings don't seem to matter a lot when it comes to engines, because often in TCEC you see engines rated at 3k doing very questionable moves, that no human rated at 2000 would do.
If ELO-3000 engines did the same moves as ELO-2000 humans, they would not be ELO-3000 engines. ;)

This topic has been archived and can no longer be replied to.