Forums

Did I break the Komodo analysis?

Sort:
plux

In analysis of a game I just played (https://www.chess.com/game/live/46875601503) -- at move 31 for white, the analysis board displays this for the various lines it is looking at:

 

The "top line" or recommended move is 31. Rd1. This is at Komodo 2901.00 depth 28. It's clearly the best move according to the engine, +25 in valuation, 4.3 points above the next best option.

 

I make the move 31.Rd1 as recommended, and then the analysis turns in to this:

"Rd1 is an inaccuracy" is displayed (?!) with an evaluation plummeting to  +6 with best play from Black. Note that this is also with the same engine depth (Komodo 2901.00 @ 28).

 

I suppose this could have something to do with a catastrophe the engine is able to see "just outside" of the engine depth, which becomes apparent only after making the move (ie the analysis is right "on the edge" of some calamity on the board, but it isn't seen pre-moving Rd1). The only problem with that theory is that I am finding the same move discrepancy at lower engine depths (~20 moves) as well. Which, to my simplistic mind anyway, seems to ruin that theory.

 

Does anyone have any other ideas / theories about what can cause this kind of issue? (a recommended move suddenly becoming a bad move, aside from issues with the depth of analysis (which seems not to be the situation here))?

 

Martin_Stahl

It's very likely a horizon issue. If you let both positions go a lot deeper, they may end up with similar evals. 

 

That said, I've seen similar things with local analysis before with Stockfish; there are some positions that it takes an additional ply for them to evaluate better.

plux

horizon issue, that's a great term for it.

 

If that were the case, I'd expect the marked discrepancy at d=20 to suddenly disappear at deeper levels (but the same 'error' over one half-move appears at d20 and d28). But maybe I have some assumptions re: how the engine works that are incorrect.

Martin_Stahl

It could also be do to some tree pruning that becomes obvious once stepping into the line. It's not extremely uncommon, especially for more complex positions.

 

I had a position one time that said my move was bad but another move was good. As soon as I tested that line, it was bad and some other move was good. Tried again and the same. It cycled through it for four or five moves before it said my intitial move was good. meh

plux
Martin_Stahl wrote:

It could also be do to some tree pruning that becomes obvious once stepping into the line. It's not extremely uncommon, especially for more complex positions.

 

I had a position one time that said my move was bad but another move was good. As soon as I tested that line, it was bad and some other move was good. Tried again and the same. It cycled through it for four otmr five moves before it said my intitial move was good.

 

Ahhhh that makes sense... thanks

kalafiorczyk
plux wrote:

The only problem with that theory is that I am finding the same move discrepancy at lower engine depths (~20 moves) as well. Which, to my simplistic mind anyway, seems to ruin that theory.

Yeah, you had a good sense that your idea is simplistic. All modern chess engines use both depth and seldepth (selective depth). At the same time, nearly all GUIs hide the seldepth for sake of simplicity and to confuse the enemy.

This is explained more thoroughly at https://www.chessprogramming.org/Depth .

If you want to know what the particular engine really thinks about your moves use the analysis tool backwards: jump to the last move and then step back one by one to the beginning. It endows the engine with prescience, an ability to see the future in the game you are analyzing.

Also be aware of the various caches that temporarily store the move evaluations, so you cannot willy-nilly step back and forth because it affects the hash table (transposition table) that is part of the engine.

Finally, if you are using web-browser-based chess engines be sure to disable browser features like "Startup boost", "Continue running in the background" and similar ones and always open the analysis page in a fresh copy of the browser after really quitting the previous one. Otherwise, you will constantly be getting bitten by the caching effects mentioned above.