I am aware of MuZero, but I'm not convinced it's able to beat the latest Stockfish. Stockfish continues to dominate TCEC even after the advent of more advanced MCTS searchers.
But MCTS is on my map as something to look into more deeply.
- TCEC (Top Chess Engine Championship) is a computer chess tournament organized and maintained by Chessdom.
___
My understanding is MuZero doesn't compete in TCEC, but might be wrong. Given Stockfish is open source and open source versions of MuZero exist, it would be easy enough to test the current version of Stockfish again; honestly surprised if it’s not already done, that it is not, given MuZero already beat it.
MuZero doesn't, but from what I know MuZero was only on par with AlphaZero, and while AlphaZero beat Stockfish initially, Stockfish has since caught up and continues to beat LeelaZero, an OSS implementation of AlphaZero in TCEC.
My understanding is that MuZero beats AlphaZero, but not by a significant amount. Even if Stockfish is better, to me at this point feels needless, given it’s highly unlikely a human over significant volume of matches would have any hope of beating any of them. It is possible that new approaches for masters might be found, but for average player, guessing it’s highly unlikely that anything new might be uncovered. Am I missing something?
You're not missing something. I just wanna see how far we can take this thing, you know? 4000 ELO seems within reach now. What about 5000, 6000 etc?
Computer chess is essentially a sport unto itself at this point, beating humans is sort of uninteresting, though I'm sure grandmasters will always find uses for ever stronger engines. They're mainly used to study openings, and since analysing from the start of the game is extremely complex, stronger engines are always helpful.
As you likely know, ELO has no maximum upper limit and chess has 10^120 possible matches — so in theory maximum ELO for chess has a long way to go, assuming there is even one engine that keeps improving over time to beat its prior version. Huge aspect of performance is tied to availability of compute; as such, my understanding is that if quantum computers ever became mainstream, even a “dumb” algorithm using true quantum computer would beat non-quantum algorithm over significant number of matches and the “ELO race” would then only be define by progress by the non-quantum algorithms against themselves. Basically, at this point, it’s a race that’s already theoretically over; might be wrong though.
Actually Elo does hit a maximum, because a perfect player doesn't beat an imperfect player 100% of the time: sometimes an imperfect player gets lucky and plays a perfect game.
A number of people have made conjectures on the Elo of perfect chess play, using extrapolations from data how chess programs scale. I'm not sure what the latest analysis is (things may have changed with the strength of neural network chess), but iirc they usually estimate something like 1000 Elo over current programs.
ELO does not have a maximum limit, chess does; already said this in my comment prior to your response. Also clearly stated “over significant number of matches” - not just a single match.
Perfect chess play would not set the maximum ELO, that’s not how ELO works. ELO simply ranks players that play. If a prefect chess engine existed and continued to play and there was at least one other chess engine that improved by beating its prior best version, its ELO and the prefect players ELO would continue to rise. Long, long way to go to fill the ELO ratings between current state of the art and the prefect all knowing player that’s aware of all 10^120 possible matches.
But MCTS is on my map as something to look into more deeply.