Komodo 12 with AlphaZero techniques


Back to forum

Thibault de Vassal    (2018-05-29)
Komodo 12 with AlphaZero techniques

Hi all! As you may know, Chessbase chess engine Komodo 12 is now promoted as having a Monte Carlo version, looking like AlphaZero algorithms... It would play a more human-like style. But does it really make a difference on a "normal" computer? It would be about 30 elo points stronger than Komodo 11 (quite a good improvement anyway)

Any tests or thoughts?

Peter W. Anderson    (2018-06-01 18:36:45)
Komodo 12 with AlphaZero techniques

Well I am a bit sceptical Thib.

Yes it is using Monte Carlo, but it is not using a neural network which is what made AlphaZero interesting (and hopefully will eventually make Leela really special).

The Monte Carlo version of Komodo is a lot weaker than Komodo 11 at the moment and time will tell if the Monte Carlo version will ever be as strong as alpha-beta searching, but my guess is that for normal engines it will not be as strong.

As an aside, the claim of a 30 Elo improvement from version 11 to 12 of normal Komodeo may be a bit extravagent judging by the ccrl rating lists.

Thibault de Vassal    (2018-06-01 23:04:32)
Komodo 12 with AlphaZero techniques

To say it all, I am a bit skeptical too and I agree with your opinion on Monte Carlo vs. Alpha-Beta in chess (but in certain positions types, maybe). CEGT rating lists seem to confirm that 30 elo points may be too optimistic as well.

Peter W. Anderson    (2018-06-02 09:21:07)
Komodo 12 with AlphaZero techniques

My gut feeling is that Monte Carlo is most applicable to the opening, but I am not sure that is better than alpha-beta even there.

By the way, in my judgement Leela is already very strong in the opening. If I was still playing correspondence I would be using Leela to help prepare openings.