Is Scorpio stronger than AlphaZero?

[Madeleine Birchfield]

It is now widespread knowledge that engines such as Stockfish, Leela, and Dragon are hundreds of elo stronger than the AlphaZero which destroyed Stockfish 9 in 2017. But what about Daniel Shawul's Scorpio, the sixth best engine currently at TCEC? Unfortunately, rhe most recent versions of Scorpio do not seem to be tested by the rating lists such as CCRL, so one cannot simply compare Scorpio's elo with AlphaZero's estimated elo. However, one could perform the same experiment as the Deepmind team did in AlphaZero vs Stockfish 9 with same strength hardware, only this time replacing AlphaZero with Scorpio, and comparing the results of Scorpio and AlphaZero against Stockfish 9.

[Daniel Shawul]

None of the engine's you mentioned come close to A0, because that is a real breakthrough. You can re-implement its algorithm and improve up on it and make it stronger than A0 like Stockfish/Leela is right now, but it never compares to A0 IMO. To venture where no has gone before is what matters the most.

Anyway if you really really want to know Scorpio strength, take a look at CEGT or fastgm tests, CCRL seems to have problems installing engines that require a bit more than extracting an exe. Also May I suggest you take a step back of from starting so many posts in talkchess -- I certainly didn't ask for Scorpio to be compared to A0.

[connor_mcmonigle]

...
Also May I suggest you take a step back of from starting so many posts in talkchess -- I certainly didn't ask for Scorpio to be compared to A0.

+1

[AndrewGrant]

...
Also May I suggest you take a step back of from starting so many posts in talkchess -- I certainly didn't ask for Scorpio to be compared to A0.

+1

+2

[Madeleine Birchfield]

None of the engine's you mentioned come close to A0, because that is a real breakthrough. You can re-implement its algorithm and improve up on it and make it stronger than A0 like Stockfish/Leela is right now, but it never compares to A0 IMO. To venture where no has gone before is what matters the most.

I disagree with this to be honest. Deep reinforcement learning has been around in computer chess since Giraffe by Matthew Tan in 2016, who eventually joined the Deepmind team to create AlphaGo, AlphaGo Zero, and AlphaZero. MCTS has been around since 2007. The combination of MCTS search with deep reinforcement learning and neural nets for evaluation, pruning, and move ordering already came in 2016 with AlphaGo, and fused policy-value networks and training from zero came in 2017 with AlphaGo Zero.

The go engine Leela Zero was started in 2017 by Gian-Carlo Pascutto to prove the results of AlphaGo Zero worked. Leela Chess Zero was forked from the Leela Zero project later thst year by Gary Linscott to see if the techniques of AlphaGo Zero and Leela Zero worked in chess. It just so happened that around the same time, Deepmind released the AlphaZero preprint which generated a lot of controversey over the AlphaZero-Stockfish games, and the Leela project gained a lot of publicity from outsiders who wanted Leela to replicate the controversial AlphaZero results.

AlphaZero was an incredible breakthrough for general game playing algorithms, showing that one algorithm could be used for multiple games to reach state-of-the-art performance at the time, but the real breakthrough in the core algorithms used in Leela and other GPU based engines, such as deep reinforcement learning, was already in Giraffe, AlphaGo, and finally AlphaGo Zero, from where the Leela Chees Zero project was derived.

Anyway if you really really want to know Scorpio strength, take a look at CEGT or fastgm tests, CCRL seems to have problems installing engines that require a bit more than extracting an exe.

Thanks, I was wondering why CCRL stopped testing Leela and Stoofvlees after 2018/early 2019 and never tested Scorpio. I'll have a look through the other rating lists again; for FGRL I didn't remember seeing Scorpio in any of the usual rating lists such as 60s+0.6s and 10m+6s, and the same with CGET in 4/40 for example, but maybe things have changed.

Also May I suggest you take a step back of from starting so many posts in talkchess -- I certainly didn't ask for Scorpio to be compared to A0.

Note taken. But I was curious, as it did seem that Scorpio is getting close to AlphaZero level.

[chrisw]

...
Also May I suggest you take a step back of from starting so many posts in talkchess -- I certainly didn't ask for Scorpio to be compared to A0.

+1

+2

+3

[Michel]

...
Also May I suggest you take a step back of from starting so many posts in talkchess -- I certainly didn't ask for Scorpio to be compared to A0.

+1

+2

+3

+4

What is the point of trying to stir up all these contentious issues...? It seems very close to trolling.

[chrisw]

...
Also May I suggest you take a step back of from starting so many posts in talkchess -- I certainly didn't ask for Scorpio to be compared to A0.

+1

+2

+3

+4

What is the point of trying to stir up all these contentious issues...? It seems very close to trolling.

Arguably, site traffic. Else you would have to assume that someone with intimate knowledge of just about every piece of technical and personal gossip over the last twenty years would have remained entirely silent during those twenty years and then suddenly is pouring out every possible contentious and potentially argumentative point ever raised in all those twenty years, with new contentious threads on a daily basis.

[Madeleine Birchfield]

All right, conversation done. Could the mods close this thread? Thanks.

[hgm]

[Moderation] Thread locked at request of OP