Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Rybka Discussion / Rybka 3 puzzle
- - By Dragon Mist (****) Date 2008-08-09 06:52
Rybka 3 mp x64 on infinite, WIN XP PRO x64, CB8 GUI, 1GB for hash, q6600 @ 3.0, default settings except Nalimov usage to Normaly, and MultiCPv= 1000. Output:

..............
43...Qg8 44.Re1 b3 45.Qxc3
  µ  (-0.96)   Depth: 18   00:30:42  420851kN, tb=39
43...Qg8
  µ  (-1.16)   Depth: 19   00:41:14  580575kN, tb=44
43...Qf7 44.Re1 b3 45.Qxc3
  µ  (-1.19)   Depth: 19   01:19:32  1154675kN, tb=134
43...Qf7 44.Re1 b3 45.Qxc3 Rxa4
  µ  (-1.19)   Depth: 20   01:49:52  1610549kN, tb=149
43...Qf7
  µ  (-1.39)   Depth: 21   04:49:13  160019kN*, tb=1513
43...Qg8** 44.Re1 b3 45.Qxc3
  µ  (-1.22)   Depth: 21   05:01:31  294324kN, tb=10837***

and is now showing (after 8 hours) that she is thinking of 43...Qg8 (1/30) at depth 22.

Notes:
* don't mind this, my system is known to reset the total node count after a 2-3 hours; it was the same with R232a
** so she chooses a move with lower evaluation (-1.22 over -1.39)???
*** this is also weird, watch the rate tbs are accessed, it is almost impossible that she accessed so many of them (from 1513 to 10837) in only 12 minutes or so (from 4h49m13s to 5h01m31s)??!
Parent - By Arkansaw (***) Date 2008-08-09 07:15
if 10837 is the number of positions in tablebase, I would think it is possible
Parent - - By Lukas Cimiotti (Bronze) Date 2008-08-09 07:34
What's the Problem? Rybka found out at depth 21 that Qf7 isn't as good as she first thought - so she switched to Qg8.
The number of tablebases accesses is perfectly normal. If I use Deep Shredder and tablebases are on a SSD, I get up to 10000 TBAs per second.
Parent - - By Dragon Mist (****) Date 2008-08-09 08:16
Don't you see? R3 finishes depth 20 with 43...Qf7 as best move for black, with 1.19 evaluation for the black side. Then, at depth 21, she sees that the same 43...Qf7 is worth even more, 1.39. Then, as she goes through the rest of the moves (still at depth 21!), she chooses 43...Dg8, although this move brings "only" 1.22. WHY would engine choose move with lower evaluation???
As for tablebases, there is a nonproportional access rate between reported 4h49m13s and 5h1m31s, which is about 9000 tb positions for only 12 minutes, compared to previously reported rates. I expect this to slightly increase (access rate that is) as search gets deeper, but this is too huge a jump. This itself is not wierd, it is unexpected when you check access rates previously! This is a 22-men positon on board, and there is no logic for sudden speed up in tb accessing.
Parent - - By Vempele (Silver) Date 2008-08-09 08:23

> WHY would engine choose move with lower evaluation???


Unless your GUI reports scores for the side-to-move, this is a bug. :)

> As for tablebases, there is a nonproportional access rate between reported 4h49m13s and 5h1m31s, which is about 9000 tb positions for only 12 minutes, compared to previously reported rates.


Qg8 might lead to TB positions at just the right moment (i.e. the lowest depth at which TBs are accessed).
Parent - - By Dragon Mist (****) Date 2008-08-09 08:27
"Unless your GUI reports scores for the side-to-move, this is a bug."

My GUI reports + for white, and - for black. It's standard, though old, CB8 that came with Shredder9. So obviously, something is wrong here.

"Qg8 might lead to TB positions at just the right moment (i.e. the lowest depth at which TBs are accessed)."

It might, but look at previous depths, where Qg8 emerged on top, only a few tb access. Can't change all that dramaticaly for this move in a matter of 1 or 2 plies?!
Parent - - By Vempele (Silver) Date 2008-08-09 08:52

> Can't change all that dramaticaly for this move in a matter of 1 or 2 plies?!


Of course it can. If the positions are now reached at the cut-off depth, the depths would've been below the cut-off depth in previous iterations (=no TB accesses).
Parent - - By Dragon Mist (****) Date 2008-08-09 09:02
Not quite sure I understand what you mean. If there's a cut-off depth for accessing TBs, how come she accessed it first time at depth 14 and after only 2 minutes search?

Here's the depths, times and Tbs:

d14   00:02:07    tb=1
d15   00:02:07    tb=1
d15   00:02:07    tb=1
d15   00:02:55    tb=2
d16   00:03:34    tb=2
d17   00:12:40    tb=7
d18   00:30:42    tb=39
d19   00:41:14    tb=44
d19   01:19:32    tb=134
d20   01:49:52    tb=149
d21   04:49:13    tb=1513
d21   05:01:31    tb=10837
Parent - By Vempele (Silver) Date 2008-08-09 09:07

> how come she accessed it first time at depth 14 and after only 2 minutes search?


Before then, all lines that reached the TBs did so at depths lower than the cut-off depth.
Parent - - By Lukas Cimiotti (Bronze) Date 2008-08-09 08:36
>Then, as she goes through the rest of the moves (still at depth 21!), she chooses 43...Dg8
No - she doesn't go through the rest of the moves, but finds something that makes her see that Qf7 isn't that good as she thought.
To verify that you might make an anylysis of Qf7 later.
Parent - By Dragon Mist (****) Date 2008-08-09 08:49
Oh come on, how many times you see this?? For over 20 years in computer chess and million of games and engines watched, I've seen something like this maybe one or twice, and with new, rude, unpolished engines. So what you're telling is the following:
after finishing depth 20, Rybka displays following as the best line/move:
43...Qf7 44.Re1 b3 45.Qxc3 Rxa4
  µ  (-1.19)   Depth: 20   01:49:52  1610549kN, tb=149

Now, she moves to depth 21. Obviously, the first considered move is 43...Qf7 (1/30). She takes some time, and after a couple of hours, she displays following:
43...Qf7
  µ  (-1.39)   Depth: 21   04:49:13  160019kN, tb=1513
which is supposed to mean that she found that her best move so far is even better by 20 cp points.

Now she moves on with the rest of the candidate moves (to see if there's even better move), and after only 12 minutes is able to tell
a) oh wait, there's a problem with 43...Qf7 (even though she was considering only 43...Qg8 at this point, and just spent 2-3 hours on 43...Qf7)!
b) doesn't display what is the problem (she might have displayed something like 43...Qf7 with say -0.98 evaluation, but this is irrational, usually when things go up (fail highs), things go up even more later at the same ply)
c) changes to 43...Qg8.

... and now as I see she refuses to display anything at depth 22, althoug she moved now to 43... Qf7 (2/30) at depth 22.
Parent - - By Vempele (Silver) Date 2008-08-09 08:55

> No - she doesn't go through the rest of the moves, but finds something that makes her see that Qf7 isn't that good as she thought.


That's not how it works. I'm pretty sure Vas would report another Qf7 PV with a lower score (probably as upperbound) if this was intentional behaviour.
Parent - By Dragon Mist (****) Date 2008-08-09 09:03
My point too!
Parent - - By George Tsavdaris (****) Date 2008-08-09 08:41 Edited 2008-08-09 08:43

>Don't you see? R3 finishes depth 20 with 43...Qf7 as best move for black, with 1.19 evaluation for the black side. Then, at depth 21, she sees that >the same 43...Qf7 is worth even more, 1.39. Then, as she goes through the rest of the moves (still at depth 21!), she chooses 43...Dg8, although >this move brings "only" 1.22. WHY would engine choose move with lower evaluation???


D20 Finishes the search and says Qf7 is the best move with -1.19 score.
D21 Starts thinking about Qf7 and thinks it is -1.39. Prints it on the screen. Continues thinking on it(while you don't see it since it doesn't report it--it doesn't print it) and finds that Qf7 is actually worth less than -1.22. Starts thinking on Qg8 and finds it to be -1.22 so it prints the result in the screen as the best move.
Parent - - By Dragon Mist (****) Date 2008-08-09 08:54
This sounds reasonable. Only I don't see why such unusal choice of what to be shown on display.
Parent - - By George Tsavdaris (****) Date 2008-08-09 10:29 Edited 2008-08-09 10:33
Yes it seems the way i described it before it is the way Rybka 3 reports its PV(see below).
In the current GUIs, because it may be related to the new Chessbase GUI that will come with Rybka 3. It may have another behavior in the new  Chessbase Rybka 3 GUI as i've seen a new pane that will report Rybka's evaluation in some other way.

Look at this position:
r5k1/pp2p2p/2nq2pB/2pNpb2/2P3P1/3r3P/PP4Q1/R4RK1 b - - 0 1


Analysis by Rybka 3 1-cpu 32-bit :
22...Bf5-e6 23.Qg2-f2 Kg8-h8 24.Ra1-d1 Rd3xd1 25.Rf1xd1 Nc6-d4 26.Nd5-e3 Ra8-d8 27.b2-b3 b7-b6
  -+  (-1.66)   Depth: 8   00:00:01  15kN
22...Bf5-e6 23.Qg2-f2 Kg8-h8 24.Ra1-d1 Rd3xd1 25.Rf1xd1 Nc6-d4 26.Nd5-e3 Ra8-d8 27.b2-b3 b7-b6
  -+  (-1.66)   Depth: 9   00:00:01  30kN
22...Bf5-e6 23.Qg2-f2 Kg8-h8 24.Ra1-d1 Rd3xd1 25.Rf1xd1 Nc6-d4 26.Nd5-e3 Ra8-d8 27.b2-b3 b7-b6 28.Rd1-f1
  -+  (-1.61)   Depth: 10   00:00:02  49kN
22...Bf5-e6 23.Qg2-f2 Kg8-h8 24.Ra1-d1 Rd3xd1 25.Rf1xd1 Nc6-d4 26.Nd5-e3 Ra8-d8 27.b2-b3 b7-b6 28.Rd1-f1
  -+  (-1.61)   Depth: 11   00:00:03  70kN
22...g6-g5 23.Bh6xg5 Bf5-e6 24.Qg2-e4 Rd3xd5 25.c4xd5 Be6xd5 26.Qe4-e3 Nc6-d4 27.Bg5-h4 h7-h6 28.Rf1-f2 e5-e4 29.Bh4-g3 e7-e5 30.Ra1-d1
  =/+  (-0.39)   Depth: 12   00:00:59  1564kN
22...Nc6-d4 23.g4xf5 Rd3-g3 24.Qg2xg3 Nd4-e2+ 25.Kg1-g2 Ne2xg3 26.Kg2xg3 g6xf5 27.Bh6-g5 f5-f4+ 28.Kg3-h4 Kg8-f7 29.Rf1-d1
  =/+  (-0.67)   Depth: 12   00:01:31  2622kN
22...g6-g5 23.Bh6xg5
  =/+  (-0.39)   Depth: 13   00:02:42  4695kN
22...g6-g5 23.Bh6xg5 Bf5-g6 24.Ra1-d1 Rd3-d4 25.Qg2-f2 b7-b5 26.b2-b3 Ra8-b8 27.Qf2-h4 Rb8-e8 28.Qh4-f2
  =/+  (-0.29)   Depth: 14   00:03:20  5875kN

You will notice that Rybka avoids giving too much information about fail low and fail high.
At depth 11 it gives Be6 as -1.61 while at depth 12 and without any apparent reason(to show that Be6 is bad) it gives g5 as better with -0.39.
Obviously Rybka found at depth 12 that Be6 is bad and "below" -0.39(actually above -0.39 since we are speaking about negative numbers), but didn't report it, it didn't print it on the screen.
Parent - By Dragon Mist (****) Date 2008-08-09 11:10
Thanks George for this example. It seems that Rybka 3 brings this unusal (for me at least) behaviour. Probably not to be mistaken for a bug. Ad acta, as far as I'm concerned. Not that I like it.
Parent - - By George Tsavdaris (****) Date 2008-08-09 08:37

>What's the Problem? Rybka found out at depth 21 that Qf7 isn't as good as she first thought - so she switched to Qg8.


Exactly, Rybka 3 found that Qf7 move on ply 21 has an evaluation of less than -1.22 so it preferred Qg8 that has -1.22.

Perhaps Qf7 had a score of exactly -1.22 and according to Vas for older Rybka's(i don't know if this still exists) Rybka 3 have chosen at random the move between Qf7 and Qg8 that had the same score.

As for tbs accesses i don't see any problem. Obviously tbs accesses have also been reset so that 1513 at ply 21 was actually much more. So the 10837 is normal.....
Parent - By Dragon Mist (****) Date 2008-08-09 09:04 Edited 2008-08-09 09:07
"As for tbs accesses i don't see any problem. Obviously tbs accesses have also been reset so that 1513 at ply 21 was actually much more. So the 10837 is normal..... "

This crossed my mind, but still the discrepancy is huge, and I've never seen TB count reset so far, while I see total node count reset EACH time (somewhere between 2h30m and 3h).
Parent - - By Dragon Mist (****) Date 2008-08-09 08:23
Update on this situation. It gets more unusal! I let the machine go on, it is now over 10 hours on infinite. Apparently, she now shows to be considering 43...Qf7 (2/30) at depth 22, but there's no output for 43...Qg8 at same (that was the move considered for the last couple of hours as 43...Qg8 (1/30))!! Last shown line is still this one:

43...Qg8 44.Re1 b3 45.Qxc3
  µ  (-1.22)   Depth: 21   05:01:31  294324kN, tb=10837
Parent - - By Dragon Mist (****) Date 2008-08-09 08:53
More update. She browsed through all the moves quickly on depth 22 (having taken 4 hours on 43...Qg8 (1/30), and about half an hour on 43...Qf7 (2/30), all without displaying anything new in the evaluation pane, and is now back at depth 22 with 43...Qg8 (1/30), meaning she starts depth 22 all over again. WTF??

Now this behaviour I've seen sometimes, at Naum I think, and maybe Zappa (or was it Crafty?).
Parent - - By Vempele (Silver) Date 2008-08-09 09:24

> WTF??


The earlier PV failed low, she tried to find another move (it'd have saved a lot of time) but couldn't, and returned to searching the PV move with a widened window.

I can't think of any other reason this would happen.
Parent - - By Dragon Mist (****) Date 2008-08-09 09:26
Yes, you're probably right. Still, no way to communicate with user like this. :-)
Anyway, thanks Vempele, Kullberg and George for putting up with me on this one. :-)
And Arkansaw, of course. :-)
Parent - - By Nelson Hernandez (Gold) Date 2008-08-09 10:40
Actually your observations were very informative.  Clearly this engine displays sequential data differently from what we've all come to expect.
Parent - By Dragon Mist (****) Date 2008-08-09 11:15 Edited 2008-08-09 11:35
Let's finish this on the bright side: so, not only does new Rybka 3 rarely reveals what she expects the other side to move, but, on occasion, she won't tell even what she will play! :-)

Edit: now I expect Vas to come and say "We'll add this to FAQ." :-)
Parent - - By Vasik Rajlich (Silver) Date 2008-08-09 16:21
43...Qf7 44.Re1 b3 45.Qxc3 Rxa4
  µ  (-1.19)   Depth: 20   01:49:52  1610549kN, tb=149
43...Qf7
  µ  (-1.39)   Depth: 21   04:49:13  160019kN*, tb=1513
43...Qg8** 44.Re1 b3 45.Qxc3
  µ  (-1.22)   Depth: 21   05:01:31  294324kN, tb=10837***

What apparently happened here is that 43. .. Qf7 first failed high at depth = 21 (Rybka's window is 20 centipawns) and then during the research failed low (which is not reported). This is rare but can happen.

Vas
Parent - By Hamlet (**) Date 2008-08-09 18:06
I have seen this behauviour too. it seems to happen sometimes.
Up Topic Rybka Support & Discussion / Rybka Discussion / Rybka 3 puzzle

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill