Komodo TCEC is doing really well so far against Stockfish

[JJJ]

All in the title. I m a little bit surprised here. I didn't expect Komodo doing so good, sample is yet small, but still, usually the lead of Stockfish is slighty more. So , is Komodo making good progress lately after all ?

[Gary Internet]

There's an update from Larry from 2 weeks ago - http://talkchess.com/forum3/viewtopic.p ... 26#p776469

I've also read in the TCEC chat that new compiles of Komodo are even fast still, and there's some other improvements that I'm not technical enough to understand. I think it's good that Komodo is improving. Houdini 6.03 is now a benchmark engine, having been in used in 4 consecutive seasons of TCEC, and having lost badly to SF in Season 11 and in CCC Rapid Rumble Final as well. It doesn't look like Houdini will be updated any time soon. Without Komodo, the only thing that might pose a threat to SF is Lc0. Everything else is much too weak.

[Mike S.]

It has certainly helped much that the startup performance problem has been fixed. But also, the intermediate result lets us predict a 12-2 or 12-3 by wins, only. Anyway, one thing is clear: This Komodo cannot be steamrolled!

[Eelco de Groot]

It seems to me that Houdini in a direct confrontation will get beaten on speed. Even on TCEC where pure speed should not matter it will get beaten because of a poorer branching factor. Komodo is a much tougher opponent. It is a huge oversimplification, but it has better knowledge.

[sovaz1997]

It seems to me that Houdini in a direct confrontation will get beaten on speed. Even on TCEC where pure speed should not matter it will get beaten because of a poorer branching factor. Komodo is a much tougher opponent. It is a huge oversimplification, but it has better knowledge.

The strength of engine doesn't depend on the branching factor, but depends on the quality of cuts and extensions on average

[yurikvelo]

because of a poorer branching factor

there is no such thing as branching factor at least for last decade, because there is no such thing anymore as depth.
Despite "depth" variable is still reported as part of UCI protocol, it is meaningless. It has no meaning even within the same engine

[Laskos]

because of a poorer branching factor

there is no such thing as branching factor at least for last decade, because there is no such thing anymore as depth.
Despite "depth" variable is still reported as part of UCI protocol, it is meaningless. It has no meaning even within the same engine

I think it can still be called "iteration". Also, despite sparse exploring due to reductions and pruning, the approximate Elo value of a "ply" stayed almost the same as with almost pure AB engines of 1980s. So, in grand-picture, the meaningless "depth" of 30 of SF now can broadly be directly compared to meaningful depth of 7 of Mephisto.
SF, Komodo et al. are that strong, they have excellent tree reduction techniques (aside much better or faster eval too).

[yurikvelo]

Komodo, Houdini and Stockfish report very different Depth variable, but it doesn't mean that their search tree cones are very different in shape/length/width/branching.

[Laskos]

Komodo, Houdini and Stockfish report very different Depth variable, but it doesn't mean that their search tree cones are very different in shape/length/width/branching.

It can be indirectly checked. Say by long forced lines or by Elo value of a ply. Sure, it would be just a guess, as different reductions and extensions can behave differently, and the overall shape of the tree can be obscured. Still, I remember doing some experiments several years ago, and both SF and Komodo showed approximately the same share of deepening and widening of the tree (about half deepening and half widening). I mean on single core, SMP distorts the shape of the tree again.

[yurikvelo]

During TCEC Season 12 - Superfinal, Komodo reported 5-10 less plies, even if total nodes searched and TB hits were higher
This doesn't mean that:
- Komodo has wider branching factor
- Komodo is weaker
- Komodo has less search depth
- Komodo search tree is different from SF in shape, length, width, branching

Whatever metrics is chosen for UCI variable "Depth" - it's meaningless and comparable to nothing, even between versions of the same engine it's incomparable and irrelevant.