Fixing New Analysis

This is the only thing I'm aware of.

https://www.chess.com/article/better-than-ratings-chess-com-s-new-caps-system

@Martin_Stahl thanks  

I was hoping for something more meaty, but looks like chess.com have not released their computational mathematics. Anyways I am going off topic on this thread, but I see lots of discussion about a system of scoring we technically don't really understand. Compare this to say Professor elo's rating system where their is much info on how it works.  But thank you . It would be great if chess.com maybe released more info on caps and how they mathematically arrive at the score. Because all debate by chess.com members at present is pure supposition.

@dallin I don't know are you are aware of this bug. Others have complained about it in other parts of the forums. @JDcannon admitted it was an oversight when planning the new analysis system but that it would be looked at...but nothing has been done to date.

Issue/bug: when you analyse a game all info in the self analysis section that was saved - variations, annotations, you name it, is wiped out and lost. It is replaced with the analyzed game data. JDcannon said this is because the new analysis system shares the same database as the "self analysis" section. 

Do you have plans to fix this? 

@flashlight002 the bug you mentioned is happening in limited circumstances, but is known, and a fix will be in place this week (likely tomorrow.)

Regarding processing power, and why we can run Analysis so quickly on our servers, I can disclose that our analysis system runs on a network of distributed servers. The exact number of computers changes, but usually includes several hundred physical cores of Intel Xeon Scalable processors. Individual machine specs vary, but our most popular configurations are Xeon Platinum 8124M CPU processors with 140 GB of ram, or Xeon Gold 6154 @3.7 GHz with 96 GB ram.

@dallin thanks for this very interesting info. That's hectic fire power. Now I get it how you can power through an analysis! Thank you for sharing this . I read in another forum that an article will soon be put out explaining caps and the accuracy score in more depth. Great to hear! 

I was going to mention that but you have already run across it. I saw it yesterday and didn't think to post about it immediately.

So, e5 should be the best move for white here when any other engine I ask instantly suggests d4? This move did not even appear among the five suggested moves for white one step earlier in the analysis. Am I missing something or do I missinterpret the tool?

You are "misinterpreting the tool"

The early stages of a game , when still in book openings, is not for an engine to guide anyone;  for Opening guidance use a database.

@ UWannaFork these results by the engine are nonsense results! The correct second line book moves for the Caro-Kann are 2.d4 d5. The variation 1.e4 c6 2.Nc3 d5 is ok as that is the 2 knights variation. 2.e5 in the games explorer has only ever been played 13 times!! So the engine feedback result is garbage! This is a clear error. @dallin am I missing something? 

I went and checked the moves using the analysis board now, just as a test. First I did it with 'show lines' and 'self analysis' checked, with engine time limit set to 1 minute. It returned the correct analysis (d = 22):

I then unchecked 'self analysis' so that the engine 'feedback' feature was switched on. Again the correct result was returned:

It also correctly identified the opening. So there clearly was something wrong with the engine when it did your analysis @UWannaFork.

Unless I an missing something. @dallin why did the engine return such a bizarre result when it did its analysis on @UWannaFork's game? This should not be happening. 

Good evening @Toire

@Toire that's not the point of the matter. Certainly I agree with you that in principle one should not rely on an engine when deciding on one's opening. We all have our favourite openings, an opening may work well against a certain competitor, or one may want to set the nature of the game (open, closed, semi closed) . However the engine is supposed to recognize and understand all the opening book moves. It also should, based on the game it is analysing, recognise the correct opening (it did not) and return result/ feedback consistent with the book moves for that opening (it did not). 

Evening @PawnstormPossie

You will see that the example I posted where self analysis Multipv sat at d = 22, and already gave the right answers interestingly enough.... but then again I had only entered a few plies. The feedback section was at d = 30, for the same position. I think I have also seen it go to 35 sometimes in my game analyses - but just for the opening phas- and only the first few plies. The rest of the game is all at d= 20 as you point out. 

I am inclined to agree with you however that setting the full game analysis depth at 20 is not sufficient, in my opinion. Generally I get very favourable results at d = 23, 24,25, 26 or so. 

The crux of the matter is that not all games have the same move tree complexities. As such having a 'one size fits all depth setting' to my mind will not give accurate results in the end. All my heavyweight analysis programs also have a time per move analysis setting one can change. The longer one gives the engine→ the more accurate the results. The current game analysis system does not allow for that kind of setting. The old system had time as a factor built in as a full analysis was about 7 minutes, and a quick analysis 2 minutes. 

I really wish chess.com could improve the accuracy of this analysis engine, or offer us a range of 'accuracy' options for the full game analysis. 

Given the processing power of the chess.com servers I keep on asking myself →why can't they offer a deeper scan?  

100% agree with you! 

I also suggested to@dallin that there should be a re- analyze button. Then @UWannaFork could have redone his game analysis with the newer engine analysis version. I hope they can put in this little extra feature. 

@PawnstormPossie what is the game you have posted? I am not following what you are trying to show/illustrate/explain. Is this a Stockfish vs Stockfish game? 

It's late, I am tired, maybe it's staring me in the face...but give me some more info on what you have posted . 

@PawnstormPossie ok understand now. I ran the game above through the analysis board and it did the same thing as it did to @UWannaFork. The analysis showed 2.e5 as a best move for the Caro Kann Defense, yet when one ticks "self analysis" and "show lines"  with MultiPV (d=22, engine time linit at 1 minute) it shows what should have been the correct move... 2. d4 d5. @dallin what is going on?

Here are the screenshots:

First the analysis and feedback system results:

Now with the move 1...c6 clicked and options "self analysis" and "show lines" ticked in order to have MultiPV active and showing engine best move for move 2. Correct lines for Caro Kann book moves show in line 1…with the 2 Knights Defense in line 2.... at d=22:

Good evening to you too:- I am determined not to come across as rude again, so will respectfully withdraw from this Thread.

flashlight002 Are you expecting the engine to tell you what the 2nd best move is? that's crazy man, All book moves are good/not best. 

No, the engine says 1.e4 c6 2.d4 not 2.e5 ,but I dunno what engine you are using. 

2e5 isn't bad  either      5

Am using chess com engine for analysis. I checked with other engines for reference. 2.d4 is what am seeing. I will never play 1c6. 1.c5 looks (and is) better.