Mixed up engine

@Zubbubu said in #21:
> @jomega
> Thank you for the information, but in the morass of details a simple point is not clear to me. There are two distinct evaluations: the arrow indicator and the analysis. Are you saying one is on the server , the other on the browser ?? Are they both stockfish but different depths? You're going into a lot of detail but my original question remains.

When you ask for a Lichess analysis, that is sent to fishnet. Fishnet is a network of computers some of which are owned by Lichess users who are providing cpu time on their local machines. Some of the fishnet clients may also be owned by Lichess, and hence might be referred to as "server machines". Regardless, the parameters of the analysis are the same for this type of analysis.

The other type of analysis you are looking at is done locally on your machine and in your browser.

Fishnet is using a parameter to Stockfish that tells SF to go 1.5 million nodes; approx. The depth will vary on that depending on the position.

Lichess use of Stockfish in the browser is using the SF depth control. What the actual depth number is will depend on the user's browser. The number of nodes in that case is not specified. However, as revoff said, modern browsers will typically surpass the current fishnet number of nodes.

The actual version of Stockfish used by fishnet clients for standard chess is Stockfish 14 + NNUE.

The version of Stockfish used in the browser depends on the browser. With a version of the browser that supports it, Stockfish 14 is used, and the user has control over whether the NNUE is used or not.

@jomega
Ok, let me see if I understand. "When you ask for a Lichess analysis" you mean the one that tells you "blunder, mistake, inaccuracy", and "The other type of analysis you are looking at" is the arrow indicators?

That raises further questions...

@Zubbubu said in #23:
> @jomega
> Ok, let me see if I understand. "When you ask for a Lichess analysis" you mean the one that tells you "blunder, mistake, inaccuracy", and "The other type of analysis you are looking at" is the arrow indicators?
>
> That raises further questions...

Yes.

Things are actually more complex than I said, however I was trying to give a simple explanation.
The bottom line is that even if you get SF 14 + NNUE for both of these; the results may not be the same. And since the user has control of the SF parameters, somewhat, in the brower, it may not match the result that other people get.

@jomega Interesting, thanks. So when there is a discrepancy, one can pretty much assume that the local recommendation (the arrows) is inferior to the Fishnet evaluation.

@Zubbubu said in #25:
> @jomega Interesting, thanks. So when there is a discrepancy, one can pretty much assume that the local recommendation (the arrows) is inferior to the Fishnet evaluation.

No. revoff was saying local is usually better.

revoff:
"...indeed analysis in the browser surpasses it quite quickly, when modern web assembly features are supported by the browser."

The "it" in "surpasses it" is referring to fishnet analysis.

Better than all this is to download Stockfish to your local machine, and then you'll have control over the analysis yourself. If you learn the UCI protocol, then you can specify the number of nodes for Stockfish to use even if your GUI does not support that.

- UCI Protocol.
http://wbec-ridderkerk.nl/html/UCIProtocol.html

@jomega Yes, you had said that:

"However, as revoff said, modern browsers will typically surpass the current fishnet number of nodes."

However, you said that the local ond is the arrow indicator. I began this thread by providing an example of where the arrow indicated a move that the Fushnet evaluator deemed an inaccuracy. I would be surprised if a move deemed an inaccuracy is actually the best move (although the engine in my brain also chose that move!)

So do you think the alleged inaccuracy could be the best move, or do you think maybe my browser is so limited that the engine generates recommendations that are inaccurate?

@Zubbubu said in #27:
> @jomega Yes, you had said that:
>
> "However, as revoff said, modern browsers will typically surpass the current fishnet number of nodes."
>
> However, you said that the local ond is the arrow indicator. I began this thread by providing an example of where the arrow indicated a move that the Fushnet evaluator deemed an inaccuracy. I would be surprised if a move deemed an inaccuracy is actually the best move (although the engine in my brain also chose that move!)
>
> So do you think the alleged inaccuracy could be the best move, or do you think maybe my browser is so limited that the engine generates recommendations that are inaccurate?

Fishnet said that move was inaccurate. Your browser SF said not so. In my browser it has the top 3 moves all +0.8, and Ne5 is one of those. So Ne5 could be one of the best moves.

#28
Actually, Stockfish and Fishnet do not have the words inaccuracy/mistake/blunder in them. Those are words that the Lichess code puts in after getting the results from Fishnet analysis. Lichess code determines how much of a point swing gets marked and mentioned in the analysis. Also, whether to say anything at all depending on winning chances.

See
- How the game analysis works.
lichess.org/blog/WFvLpiQAACMA8e9D/learn-from-your-mistakes

Isn't there also a difference between post-game browser requested analysis (whole game lichess mediated analysis via SF calls), and per position manual (toggle above move list in analysis page) analysis? (some of it, visible in the displayed parameters, but some discrepancies behind the curtain still manifested). The manual toggle per position SF call is affected by the user preferences, among others the PV size. So I am not sure what the arrows listen to in that 2 or 3 prong contingency (whole game, position, and local versus fishnet). Could someone make a simple clarification table of the differences that have an effect at the conscious user end (meaning noticing discrepancies in analysis results for same position, e.g.).?

PS: This question is for further enjoyment of the lichess platform... not for complaining and demanding... There might be a gap of documentation (rather non-unique it seems in fast paced development) between user end pragmatic (but rational, curious and analytical if not critical) and source code line reference, even if source code its genius about its designing development, high level language that makes maintenance reasonable and expanding features natural (it seemed going in that direction from the little I could fast read).

@dboing said in #30:
> Isn't there also a difference between post-game browser requested analysis (whole game lichess mediated analysis via SF calls), and per position manual (toggle above move list in analysis page) analysis? (some of it, visible in the displayed parameters, but some discrepancies behind the curtain still manifested). The manual toggle per position SF call is affected by the user preferences, among others the PV size. So I am not sure what the arrows listen to in that 2 or 3 prong contingency (whole game, position, and local versus fishnet). Could someone make a simple clarification table of the differences that have an effect at the conscious user end (meaning noticing discrepancies in analysis results for same position, e.g.).?
>
> PS: This question is for further enjoyment of the lichess platform... not for complaining and demanding... There might be a gap of documentation (rather non-unique it seems in fast paced development) between user end pragmatic (but rational, curious and analytical if not critical) and source code line reference, even if source code its genius about its designing development, high level language that makes maintenance reasonable and expanding features natural (it seemed going in that direction from the little I could fast read).

The arrows correspond directly to how many 'Multiple lines' the user has configured under the sandwich menu. The arrows correspond to the MPVs displayed above the moves.