Not logged inRybka Chess Community Forum
Up Topic Correspondence Chess / Correspondence Chess / titanium cranium Vs. Vytron (Cancelled)
1 2 Previous Next  
- - By Uly (Gold) [mx] Date 2009-07-01 00:57 Edited 2009-07-03 12:11
24h/move as a guideline but overstepping time limit doesn't mean time forfeit.

titanium cranium is restricted to only Rybka 3 under Deep Position Analysis.

Vytron is restricted to only Rybka 3 with Infinite Analysis.

*Handshake*

Your move, good luck!
Parent - - By titanium cranium (***) [us] Date 2009-07-01 01:19 Edited 2009-07-01 01:45
Ok, let's start the game. How do we post our moves and comments? Also, it might make more sense to play the first game with you playing only IA, then try interaction and IA in game 2.
Parent - - By Uly (Gold) [mx] Date 2009-07-01 01:55
I suggest putting a hard limit then, because otherwise, DPA could win just because you're letting it analyze much more time than I do.

I suggest 30/min move, so you put DPA to analyze for 30 minutes and play the best move suggested, then I put IA to analyze for 30 minutes and play the move suggested. To avoid interaction you'd start a new DPA from scratch every move, and I'd clear the hash every move.

Please look at this to see how forum games have been handled:

http://rybkaforum.net/cgi-bin/rybkaforum/board_show.pl?bid=3

I could copy and paste the analysis output of Rybka and you could copy and paste the DPA output (so people can paste it to their Fritz GUI if interested.)
Parent - - By Nelson Hernandez (Gold) [us] Date 2009-07-01 03:18
(Pssst, Vytron!  You should slaughter him!)
Parent - By Uly (Gold) [mx] Date 2009-07-01 03:45
(:wink:)
Parent - - By Banned for Life (Gold) Date 2009-07-03 21:21
Your methodology is badly flawed. DPA is obviously much less likely to miss the opponent's next move, so starting from scratch is clearly the wrong thing to do.
Parent - By Uly (Gold) [mx] Date 2009-07-04 03:19
Ok, then he can start at the point he left, and I'll keep saving and reloading the hash, it will be interesting to know if successful move prediction is going to be a big factor, like with IDEA. Thanks.
Parent - - By Dragon Mist (****) [hr] Date 2009-07-01 04:55

> interactive Infinite Analysis


What is this?? :-)
Seriously, what are you constraints in making your analysis, Vytron?
Parent - - By Uly (Gold) [mx] Date 2009-07-01 05:14
If I understood correctly, titanium cranium claims that DPA is as good as Infinite Analysis even if used interactively, he challenged me to a game and I couldn't say no! How are we going to have it is still unclear, but I'm restricted to Rybka 3 (no other engines) IA (no other analysis methods like DPA or IDEA) and I'll restrict myself to 3CPU (to approximate equal hardware.)
Parent - - By titanium cranium (***) [us] Date 2009-07-01 12:31
Actually, I'm simply interested in determining the strongest combination of game tools, e.g., Aquarium 3.1.1, Deep Rybka, human interaction, etc. But the only differences in each side here has to be DPA v IA, or it will be totally meaningless. For example, I am using Aquarium 3.1.1 and commercially available Deep Rybka 3, with no enhancements or secret add-ons.

Bad luck. E4.
Parent - - By Uly (Gold) [mx] Date 2009-07-01 12:50
1...c6
Parent - - By titanium cranium (***) [us] Date 2009-07-01 14:55 Edited 2009-07-01 14:58
We know that Aquarium indicates that black has a high success rate with c6 here, but would you actually play c6 if you HAD to win as black? In other words, I agree that black should easily draw after c6, but as for winning  . . . ?

D4.
Parent - By Uly (Gold) [mx] Date 2009-07-02 01:59
Yes, I think IA is so superior to DPA (on equal conditions, I think we agreed to let them run for 30 minutes) that the Caro-Kann is the correct choice.

2...d5

1. e4 c6 2. d4 d5
Parent - - By Dragon Mist (****) [hr] Date 2009-07-01 15:08
I am pretty sure IA will come on top. Do you people plan to post any AN of the moves? Even with delay?
Parent - - By Uly (Gold) [mx] Date 2009-07-02 01:59

> Do you people plan to post any AN of the moves?


Yes, as soon as we get out of book we're going to start posting full engine output.
Parent - - By titanium cranium (***) [us] Date 2009-07-02 12:30
Nc3. Aren't you going to post the game on the pgn above?
Parent - - By Uly (Gold) [mx] Date 2009-07-03 12:11
Nah, it's actually faster to do it each move.

3...dxe4

rnbqkbnr/pp2pppp/2p5/8/3Pp3/2N5/PPP2PPP/R1BQKBNR w KQkq -
Parent - - By titanium cranium (***) [us] Date 2009-07-03 12:52
Nxe4
Parent - - By Uly (Gold) [mx] Date 2009-07-03 13:30
4...Nd7
Parent - - By titanium cranium (***) [us] Date 2009-07-03 18:05
Nf3
Parent - - By Uly (Gold) [mx] Date 2009-07-04 07:53
5...Ngf6

r1bqkb1r/pp1npppp/2p2n2/8/3PN3/5N2/PPP2PPP/R1BQKB1R w KQkq -
Parent - By titanium cranium (***) [us] Date 2009-07-04 13:21
6. Ng3
Parent - - By Bobby C (****) Date 2009-07-04 08:01
I do not think I have seen this variation in a while without 4...Bf5. Different but very interesting. Did you decide on the CK for many of the same reasons I played it in our last match? I think it gives you the advantage here vs others like the Sicilian and Spanish. It is funny because I had won a game with the CK and now days my whole goal in life is to crush those who play it but that is in blitz games, in long games it is a very solid opening and damn good to play as black especially if you have more experience with it then your opponent. Good luck my friend :) !
Parent - - By Uly (Gold) [mx] Date 2009-07-04 08:47

> Did you decide on the CK for many of the same reasons I played it in our last match?


Yes.

> Good luck my friend :-) !


Thanks, I hope Rybka 3 doesn't let me down, as for this game it's forbidden to interact with her, but it's a win-win situation, either I win or we discover that DPA is capable of finding better (or equal) moves than IA in the same time.
Parent - By titanium cranium (***) [us] Date 2009-07-05 22:04
I use a random, subjective combination of IA, IIA and DPA in my games, so we should also play a game IIA v. DPA.
Parent - - By Uly (Gold) [mx] Date 2009-07-05 02:05
Well, I say we've had enough book, let's turn on the analysis methods. As we agreed, we'll run them for 30 minutes and play the best move chosen by the engine. Here's this move analysis:

Rybka 3 64bit 3CPU 64MB RAM
11.00   0:01   +0.32   6...h5 7.h4 Nb6 8.Be2 g6 9.O-O Bg7 10.c4 (113.343) 95
12.01   0:02   +0.26   6...h5 7.h4 Nb6 8.Be2 g6 9.O-O Bg7 10.c4 Qc7 11.b3 (245.911) 97
13.01   0:03   +0.26   6...h5 7.h4 Nb6 8.Be2 g6 9.O-O Bg7 10.c4 (350.155) 96
14.01   0:05   +0.26   6...h5 7.h4 Nb6 8.Be2 g6 9.O-O Bg7 10.c4 (518.417) 99
15.01   0:12   +0.27   6...h5 7.h4 g6 8.Bd3 Nb6 9.a4 Bg4 10.a5 Nbd5 11.O-O Bg7 12.Bd2 O-O 13.Qe2 (1.267.782) 102
16.01   1:02   +0.42   6...h5 7.h4 g6 8.Bd3 Nb6 9.a4 Bg4 10.a5 Nbd5 11.O-O Bg7 12.Bd2 O-O 13.Qe2 Qd6 14.a6 b5 (6.031.752) 99
16.02   1:54   +0.20   6...g6 7.Bd3 Bg7 8.O-O O-O 9.Qe2 Re8 10.Re1 Qb6 11.a4 c5 12.a5 Qc7 13.c3 cxd4 (12.098.193) 108
17.01   2:14   +0.20   6...g6 7.Bd3 Bg7 8.O-O O-O 9.Qe2 Re8 10.Re1 Qb6 11.a4 c5 12.a5 Qc7 13.c3 cxd4 (14.137.686) 107
18.01   3:32   +0.25   6...g6 7.Bd3 Bg7 8.O-O O-O 9.Qe2 Re8 10.Re1 e5 11.dxe5 Ng4 12.Bg5 Qb6 13.Ne4 Ndxe5 14.Nxe5 Nxe5 (22.394.310) 107
19.01   5:12   +0.25   6...g6 7.h3 Bg7 8.Bc4 O-O 9.O-O b5 10.Bd3 Nb6 11.Qe2 Be6 12.b3 Nbd5 13.Bd2 Qd6 14.Ne5 c5 (32.926.001) 107
20.01  24:06   +0.25   6...g6 7.Bc4 Bg7 8.O-O O-O 9.Re1 Nb6 10.Bb3 a5 11.a4 Nbd5 12.Qd3 Qc7 13.Bd2 b6 (157.278.188) 111
best move: g7-g6 time: 30:00.718 min  n/s: 108.608  nodes: 190.985.058

6...g6

1. e4 c6 2. d4 d5 3. Nc3 dxe4 4. Nxe4 Nd7 5. Nf3 Ngf6 6. Ng3 g6

r1bqkb1r/pp1npp1p/2p2np1/8/3P4/5NN1/PPP2PPP/R1BQKB1R w KQkq -


I'll save the hash so that higher depth is reached faster, you're free to start DPA at the point you stopped it.
Parent - - By Mark (****) [us] Date 2009-07-05 02:43
Maybe the winner of this game can take on IDeA??
Parent - By Uly (Gold) [mx] Date 2009-07-05 02:51
And the loser too, it could turn out to be Rock-Paper-Scissors!
Parent - - By titanium cranium (***) [us] Date 2009-07-05 19:11
Hmm, a potential glitch has arisen. I have around 10 games underway today using the same laptop, so I have to start, stop, replace, open, change main engine parameters, etc. for each game. I will not be able restart DPA from the same point in its analysis in our game, will I? If yes, how?
Parent - - By Uly (Gold) [mx] Date 2009-07-05 23:26
I think you could save the game and then reload it? The notation should be saved. Or since you're going to paste the notation on the forum perhaps you can copy it back and it'll work.
Parent - - By titanium cranium (***) [us] Date 2009-07-06 13:09
Do I also enable persistent hash and preserve analysis, and save the file, say, Vytron black dpa v ia?
Parent - - By Uly (Gold) [mx] Date 2009-07-06 13:59
It's either persistent hash or preserve analysis or save hash, I don't think these options get along. But sure, you can turn on anything you want (as long as you use the same settings the whole game).
Parent - - By titanium cranium (***) [us] Date 2009-07-06 14:43
Well, if you're only saving hash, perhaps it makes sense for me to do the same. Do you name the file the name of the game?
Parent - By Uly (Gold) [mx] Date 2009-07-06 14:47

> Do you name the file the name of the game?


The name doesn't matter, it can be anything. I named mine "GAME" (because it's the first time I use this feature.)
Parent - - By Banned for Life (Gold) Date 2009-07-06 14:52
- Saving hash makes great sense for IA. It makes no sense for DPA.
- Persistent hash would be good for DPA if it worked correctly without assistance. Its not clear that it will help in the current implementation (you may get stuck somewhere and waste all of your time on one position). There is also interaction between being in multi-pv mode and storing/retrieving from PH.
- In some implementations of DPA, storing the moves in a book file will guarantee their inclusion on the list of considered alternatives.
Parent - - By Uly (Gold) [mx] Date 2009-07-06 14:57
In theory PH would be excellent for DPA because by doing the MultiPV switches the program is avoiding the Persistent Hash bugs by itself, but it hasn't been tested so it would be experimental.
Parent - - By Banned for Life (Gold) Date 2009-07-06 15:03
DPA will spend a lot of time in MultiPV mode. Are moves saved and read from PH while MultiPV is enabled? If not, PH will only be used near the leaf nodes (which could still be beneficial if it doesn't get stuck somewhere).
Parent - - By Uly (Gold) [mx] Date 2009-07-06 15:08

> Are moves saved and read from PH while MultiPV is enabled?


Yes. So if you have MultiPV=5, 5 moves are stored/read to/from the P.hash. MultiPV avoids Persistent Hash bugs like wrong move ordering and shows an exact score for fail lows and fail highs faster, and switching between MultiPV lines helps with backward propagation of moves' scores.
Parent - By Banned for Life (Gold) Date 2009-07-06 15:15
OK, so it looks like appropriate settings for completely automated play would be:

IA
Persistent Hash off because of known bugs
Save Hash used after each move and restored after opponent's move
Preserve Analysis off

DPA
Persistent Hash on
Save Hash not used
Preserve Analysis off

Other open issues include DPA branching and cutoff values.
Parent - - By titanium cranium (***) [us] Date 2009-07-06 15:10
Ok, I'm not familiar with any of this, so I will simply set R3 on default settings, and retain and save nothing. Then the question becomes, what are the best settings for dpa in a 30 min/move game? Also, can I turn-on the clock in dpa?
Parent - - By Uly (Gold) [mx] Date 2009-07-06 15:48
To increase your winning chances you'd listen to Alan and use Persistent Hash, together with saving all the suggested moves to a book so they're considered by DPA on future moves.

Or we can go without storing anything, the overall game level would be lower but the point will be the same, DPA is going to be very selective so that if the best move requires high depth to see, DPA is going to prune it while IA will not, and as this happens both offensively (DPA misses playing the best move) or defensively (DPA doesn't examines at all what IA is going to actually play) the playing strength degrades with each move and DPA would get squashed. I don't think maximizing your strength would change these fundamental issues (running DPA at fast hardware is akin to running IA at slow hardware) specially since then I would be saving and loading the hash to compensate.

You still have to find some optimal settings for this experiment, but you would need to talk with someone that uses DPA effectively with a similar hardware and conditions, it's not trivial (the match is still interesting because this is all still theory, I don't see how DPA could come with a better move than a full hash loaded IA, but it could happen).
Parent - - By titanium cranium (***) [us] Date 2009-07-06 16:05
Current DPA settings:

time = 1800 sec
plus root = 1800 sec
branching 1st = 3; 2nd = 2; 3rd = 2; 4th = 1.
Branching = both
length of variation = 20
1/100 pawn = 80
cut bad lines = checked
multiple engines = unchecked
Parent - - By Uly (Gold) [mx] Date 2009-07-06 16:09
Good, now let that run for 30 minutes and copy and paste the notation into the thread (it will look like a big mess when pasted, that's normal.)
Parent - By Banned for Life (Gold) Date 2009-07-06 16:27 Edited 2009-07-06 16:31
These settings are NFG:

If you are using the CB GUI, they should be something like this:

Total Time:    Selected
Total Time:   30 min

Branching (defaults)
1: 3
2: 3
3: 3
4: 3

Branching: Both

Length of variations: 8
Evaluation window:   40
Parent - - By Banned for Life (Gold) Date 2009-07-06 16:18
DPA is going to be very selective so that if the best move requires high depth to see, DPA is going to prune it while IA will not, and as this happens both offensively (DPA misses playing the best move) or defensively (DPA doesn't examines at all what IA is going to actually play) the playing strength degrades with each move and DPA would get squashed.

This is too strong a statement. Remember that DPA forces multiple moves near the root, so that suboptimal moves are examined if they are in the top N moves at a lower depth. Some moves will be missed by DPA because they have a major shift in evaluation beyond the DPA depth horizon, but other moves that are closer in score at lower depth may end up being selected by DPA.

I think the result will be that DPA may outperform IA in quiet positions, but may miss deep tactical shots featuring "surprising moves" (both offensive and defensive as you've pointed out). In this case a surprising move would be one which goes from not being one of the top N moves to the top move at a high depth.
Parent - - By Uly (Gold) [mx] Date 2009-07-06 16:23

> but may miss deep tactical shots featuring "surprising moves" (both offensive and defensive as you've pointed out).


I think one of those is enough to make the game decisive, we will see.

Anyway, still waiting for TC's move, I refuse to play against something that looks like MultiPV=3, I need to see the notation (because TC's move should be from the line of the notation that scored best.).
Parent - By Banned for Life (Gold) Date 2009-07-06 16:28
His moves are multi-pv because he set it to take 3600 seconds on the root position... :-)
Parent - - By Banned for Life (Gold) Date 2009-07-06 16:35
I'm not sure that automated IA will outperform automated DPA intelligently applied. It would take significant resources to test this though.
Parent - - By titanium cranium (***) [us] Date 2009-07-06 16:42
You seem to know a lot more about Deep R3, DPA, IA, etc., than I know, so it would seem to make a lot more sense for you, Deleted, to take over this experiment. If I'm not using DPA at its strongest settings then this entire deal is a colossal waste of time and energy, neither of which can I spare.
Parent - - By Banned for Life (Gold) Date 2009-07-06 16:57
I will be happy to run this experiment with Vytron if he chooses. I would ask to restart the game though, because my book is actually optimized to improve performance of DPA relative to IA.
Up Topic Correspondence Chess / Correspondence Chess / titanium cranium Vs. Vytron (Cancelled)
1 2 Previous Next  

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill