Komodo vs Hikaru

Discussion of computer chess matches and engine tournaments.

Moderators: hgm, Rebel, chrisw

carldaman
Posts: 2283
Joined: Sat Jun 02, 2012 2:13 am

Re: Komodo vs Hikaru

Post by carldaman »

Does Komodo MCTS use contempt? AFAIK it normally doesn't recognize the contempt setting (unlike regular Komodo) - instead, its optimistic eval acts as some sort of built-in de facto contempt.
mjlef
Posts: 1494
Joined: Thu Mar 30, 2006 2:08 pm

Re: Komodo vs Hikaru

Post by mjlef »

carldaman wrote: Wed Jun 20, 2018 12:49 am Does Komodo MCTS use contempt? AFAIK it normally doesn't recognize the contempt setting (unlike regular Komodo) - instead, its optimistic eval acts as some sort of built-in de facto contempt.
Komodo 12, and 12.1.1 automatically switch Contempt to zero for MCTS mode. We may change this in a future release so it can be used in Odds games.
lkaufman
Posts: 5960
Joined: Sun Jan 10, 2010 6:15 am
Location: Maryland USA

Re: Komodo vs Hikaru

Post by lkaufman »

Nordlandia wrote: Tue Jun 19, 2018 6:59 pm Game 2: Knight for pawn

At move 30...

Nakamura said he was worried for little time on the clock, i understand his concern.

Shouldn't it be possible to teach Komodo to take opponents time into consideration for human odds matches.

Realistically Komodo should have continued playing on, Naka may have blundered in the time scramble.

Even if the position was worse. Naka's low time on clock make it probably equal. Approximately "Equilibrium" state of matter.

Position was better for Nakamura but time trouble seriously increase blunder ratio. Komodo may have avoided threefold repetition by exploiting Naka's bad time management.
Theoretically we should raise Contempt when the opponent is in time trouble, but given the practical problems of considering increment, whether the opponent is a machine or human, what his elo is, etc. it's not very practical for us to spend our time on this. In the game 2 we used Contempt 50, which apparently was not high enough to avoid the draw at the end. Some tests I ran indicate that the elo seems to start dropping more sharply beyond Contempt 50, and I felt that 50 was about right for the strength difference between Naka and MCTS Komodo. But probably given the fast time limit, I should have picked a higher value. To answer another question, we spend 95% or more of our time optimizing for selfplay or against other comparable engines. The main value of optimizing vs. humans is to increase the score against weak engines, and for the occasional public match like this one. Sometimes we notice something that turns out to be helpful even against top engines.
Komodo rules!
lkaufman
Posts: 5960
Joined: Sun Jan 10, 2010 6:15 am
Location: Maryland USA

Re: Komodo vs Hikaru

Post by lkaufman »

carldaman wrote: Wed Jun 20, 2018 12:49 am Does Komodo MCTS use contempt? AFAIK it normally doesn't recognize the contempt setting (unlike regular Komodo) - instead, its optimistic eval acts as some sort of built-in de facto contempt.
In early versions of Komodo MCTS the "built-in" Contempt was quite strong; now it is much milder so it starts to make sense to apply Contempt to Komodo MCTS.
Komodo rules!
Gary Internet
Posts: 60
Joined: Thu Jan 04, 2018 7:09 pm

Re: Komodo vs Hikaru

Post by Gary Internet »

https://youtu.be/2R1mIJ80CJg?t=1h51m45s - Games now uploaded to YouTube. I've set this link to take you to the start of the odds matches because watching Hikaru rip through the early levels of quite dull and too fast to follow.