Out of the Kai 450 positions, 99 not found by SF

[Uri Blass]

I'm a 2300+ level player and I have NEVER met an opponent of that level that played a game so badly! NEVER!

I'm having trouble believing that that is a "real" ICCF tournament. I've never seen players of that level play so badly!

I previously assumed a "bad" CC player would follow an engine too much, not too little, so the quality of games would never be much worse than an engine-engine game. Is that not true?

It is clearly not true.
There are a lot of bad CC players that are weaker than chess engines(assuming you give the engines some hours per move on normal hardware)
Let divide them to different types(players can be in more than one type and players may be both type 2 and type 3 and type 4 and type 5.

tyoe 0 CC players do not use engines and use only their own brain.
typel 1 CC players use engines only to verify that their move is not a significant tactical mistake.
type 2 CC players use old engines that are not the top engines to play the move and do not know that stockfish and lc0 are the best engines
type 3 CC players use chess engines but not for a lot of time and their move may be result of 1 minute analysis by the engine instead of giving the engine many hours to analyze
type 4 CC players play too many games so they practically cannot give their machine a lot of time.
type 5 CC players are not careful and sometimes make clerical errors in their games(for example playing Ned7 instead of Nfd7).
type 6 CC players trust book moves not from a good book without engine analysis to verify them and they may fall into a trap in the opening because they did not use engines to analyze.

[jp]

Let divide them to different types (players can be in more than one type and players may be both type 2 and type 3 and type 4 and type 5.

<snip>

Yes, I get that these are all logical possibilities. What I want to know is what proportion of (e.g. ICCF or any other major CC organization) players of e.g. elo around 2100-2300, 2300-2450, 2450-2600, etc. fall into each of the 6 types you list.

What I really want to know is the quality of the games you can expect from CC players in different CC elo bands.

[Zenmastur]

Let divide them to different types (players can be in more than one type and players may be both type 2 and type 3 and type 4 and type 5.

type 0 CC players do not use engines and use only their own brain.
type 1 CC players use engines only to verify that their move is not a significant tactical mistake. 
type 2 CC players use old engines that are not the top engines to play the move and do not know that stockfish and lc0 are the best engines
type 3 CC players use chess engines but not for a lot of time and their move may be result of 1 minute analysis by the engine instead of giving the engine many hours to analyze
type 4 CC players play too many games so they practically cannot give their machine a lot of time.
type 5 CC players are not careful and sometimes make clerical errors in their games(for example playing Ned7 instead of Nfd7). 
type 6 CC players trust book moves not from a good book without engine analysis to verify them and they may fall into a trap in the opening because they did not use engines to analyze.

Yes, I get that these are all logical possibilities. What I want to know is what proportion of (e.g. ICCF or any other major CC organization) players of e.g. elo around 2100-2300, 2300-2450, 2450-2600, etc. fall into each of the 6 types you list.

What I really want to know is the quality of the games you can expect from CC players in different CC elo bands.

Uri's analysis of types of players is accurate. I started at ELO 1800P in 2015, my games last about a year, and my rating has gone to 2300+, so I've played players that play terrible to those that play pretty well. Generally above 2100 the play is mostly competent, players use engines all the time with only a few that don't use computers effectively. I've only played 50 games so I'm sure there is more to it. I also play in the World Zone which may have an effect on the strength of play. I used to be in NAPZ / Zone 4.

Below 2100 you get all types of players. Even competent players that use computers effectively. They don't stay below 2100 for long. By 2300 or so most of the players that have an “issue” get left behind. Getting wins are much harder above 2300 because the number of errors in play that are serious enough to cause a loss go way down. In my case it took 3 errors to lose a game.

It's very easy to tell when you are playing better than your opponent. It's obvious from his play. It's much harder to tell if you are being out-analyzed. The only game I've lost was due to me playing a move that wasn't found anywhere in my analysis AND playing 2 moves that were transposed. I've never felt like I was being out-played/out analyzed in any of my games and I was never in any serious trouble in any other game. But I haven't played any players over 2400.

Not sure that helps you much. When I was using ICCF games as opening aids I cut off the ratings at 2200 ELO because there are still many ideas at this level and the play isn't too bad. There are still errors of course, but there weren't so many that it was annoying to use.

Regards,

Zenmastur

[jp]

Not sure that helps you much. When I was using ICCF games as opening aids I cut off the ratings at 2200 ELO because there are still many ideas at this level and the play isn't too bad. There are still errors of course, but there weren't so many that it was annoying to use.

Thanks. Any ICCF players' comments are helpful. This reminds me of discussion about the offset between different elo ratings, which it would be nice to know accurately without just guessing.

[Zenmastur]

Not sure that helps you much. When I was using ICCF games as opening aids I cut off the ratings at 2200 ELO because there are still many ideas at this level and the play isn't too bad. There are still errors of course, but there weren't so many that it was annoying to use.

Thanks. Any ICCF players' comments are helpful. This reminds me of discussion about the offset between different elo ratings, which it would be nice to know accurately without just guessing.

Yeah, CC players aren't known to be a talkative lot. They mostly never say anything during the game. You get a welcome message and end of game message and that's it. I've had exactly 4 messages during 50 games. And don't even think about asking about what engine, GUI, Hardware, EGTBs, Opening books etc. as they will never tell you.

I have on several occasions wanted to tell my opponent that further resistance was futile by posting a mate in 46 with the full line of play etc, but they have a strict code of conduct so I never have. I also have wanted to tell my opponent that the game was effectively drawn 50 moves ago so why the #%LL didn't you accept my draw offer? Some people just want to put off the inevitable for rating reasons even with draws. That's when you can get creative and find a way to force exchanges, even if they are bad for you, to get to 6-piece endgame so you can claim a draw.

I generally look at their past games looking for where they are weak, not always but most of the time. See what openings they play well and see how prone they are to errors etc. I also check their game load to see how over taxed they are. Then just play my best game. It would actually be nice to have a forum for CC players but I'm not sure how many would participate.

Regards,

Zenmastur

[jp]

What are N, P, Q, D, U, V?

They're mostly capital letters to me, but we'd likely both benefit from reading here: https://github.com/LeelaChessZero/lc0/w ... Chess-Zero

P: Neural network's raw policy output (probability this is the best move)
Q: Average expected value of all playouts for this move, a value ranging from -1 to 1.
UCB: Upper Confidence Bound, aka U. This is the part of the PUCT formula that encourages exploring moves that have not been searched much yet. See also the other half, Q.
V: Expected Value output of the NN, ranging from -1 to 1.

N, D are not there, but N in their example looks like the number of visits. Is D depth maybe?

e.g.

Code: Select all

 Move            	   Visits         Policy      	Avg. Value     	 UCB                      Raw NN Value
info string e2e4  (322 ) N:     128 (+63) (P: 10.01%) (Q:  0.05778) (U: 0.05111) (Q+U:  0.10889) (V:  0.0556)

(But their example also has extra numbers. e.g. "(322)", "(+63)".)

[jp]

I am curious, do you have some confidence that a strong Lc0 is superior "positionally" to a strong regular eval AB engine?

I'm open to the possibility, but <snip>

I'm not sure that they are really giving different evaluations and PVs most of the time, when you give a lot of computing time.

Here are two positions to see whether NN & AB engines give different evaluations.

[d]r1bq1r2/pppn2k1/2Pp1b2/3Pp3/1P2PpP1/2N2P1p/P3B3/R2QNR1K b - - 1 20

The Chessbase human games-trained NN is reported to have eval +2.02, and SF -0.83. (Nodes/Depth not known.)


[d]5rk1/NpR5/3p1q1b/3Pp2p/4Pn2/5P1b/PPQ2B1P/4N2K w - - 9 29

The Chessbase human games-trained NN is reported to have eval -1.22, and SF 0.00. (Nodes/Depth not known.)


Can someone please run these positions on Leela (& SF) for a long time to see what it says?

[zullil]

I am curious, do you have some confidence that a strong Lc0 is superior "positionally" to a strong regular eval AB engine?

I'm open to the possibility, but <snip>

I'm not sure that they are really giving different evaluations and PVs most of the time, when you give a lot of computing time.

Here are two positions to see whether NN & AB engines give different evaluations.

[d]r1bq1r2/pppn2k1/2Pp1b2/3Pp3/1P2PpP1/2N2P1p/P3B3/R2QNR1K b - - 1 20

The Chessbase human games-trained NN is reported to have eval +2.02, and SF -0.83. (Nodes/Depth not known.)

Well, Sf-dev after a short search says WHITE is +2.10

info depth 38 seldepth 62 multipv 1 score cp -209 nodes 4189306898 nps 20084892 hashfull 944 tbhits 0 time 208580 pv b7c6 d5c6 d7b6 a2a4 c8e6 a4a5 b6c8 b4b5 d8e7 e1c2 e7f7 c2b4 c8e7 b4d5 e6d5 c3d5 e7d5 e4d5 f6h4 a1a4 f8e8 d1b3 a8b8 e2d3 h4g3 a4a2 b8a8 f1b1 g7f6 b3c4 e8b8 c4e4 f6g5 a2b2 b8h8 e4c4 a8b8 b2a2 b8a8 b5b6 a7b6 a5b6

And now

info depth 45 seldepth 60 multipv 1 score cp -164 nodes 12513797185 nps 20406700 hashfull 1000 tbhits 0 time 613220 pv b7c6 d5c6 d7b6 a2a4 c8e6 a4a5 b6c4 d1b3 d6d5 a1d1 c4e3 e4d5 e6g8 e1d3 f6e7 b3b2 e7d6 c3e4 d8e7 e4d6 c7d6 b4b5 g8d5 d3b4 d5f7 b5b6 e3d1 e2d1 a7b6 a5b6 f8b8 b6b7 b8b7 c6b7 e7b7 b2d2 b7b5 d1e2 b5c5 e2d3 a8a4 f1b1 c5e3 d2e3 f4e3 h1h2 f7g6 d3g6 g7g6 b4d5 a4a3 b1e1 e3e2 e1e2 a3f3 e2e3 f3f2 h2h3 g6g5 h3g3

Someone goofed.

[jp]

Someone goofed.

I'm guessing the claimed SF eval must have been for low depths (e.g. <30 or so?).
What does Leela say? I'm now guessing this is another example of them not being all that different when you give them enough time.

[zullil]

Someone goofed.

I'm guessing the claimed SF eval must have been for low depths (e.g. <30 or so?).
What does Leela say? I'm now guessing this is another example of them not being all that different when you give them enough time.

But more important, someone misunderstood a sign. At every depth in my search, Sf-dev says WHITE is ahead.

Now at

info depth 56 seldepth 109 multipv 1 score cp -225 nodes 166645749526 nps 21408155 hashfull 1000 tbhits 0 time 7784218 pv b7c6 d5c6 d7b6 a2a4 c8e6 a4a5 b6c4 d1b3 d6d5 a1d1 c4e3 e4d5 e6g8 e1d3 f6e7 d3e5 d8d6 e5d7 d6b4 b3b4 e7b4 d7f8 g7f8 c3b5 b4a5 d1d4 e3f1 e2f1 a5b6 d4f4 f8g7 f1c4 a8d8 f4f5 g8f7 f5g5 g7h8 h1h2 a7a6 b5c3 b6e3 g5f5 f7g6 f5f6 h8g7 f6e6 e3f4 h2h3 g6f7 e6e4 f4d6 f3f4 f7g6 f4f5 d8h8 h3g2 h8h2 g2f1 g6f7 c3e2 h2h3 e2d4 h3c3 d4e6 f7e6 d5e6 g7f6 c4a6 c3c6 a6d3 c6c1 f1f2 c1c3 f2e2 c3a3 e4c4 a3b3 e2e3 b3b4 c4b4 d6b4 e3e4 b4d2 e4d5 f6e7 d5c6 d2e3 d3c4 e3f4

cp -225, so from the computer's perspective (i.e., Black's perspective) things are looking grim!