"Battle of the Goths" Championship, cycle #1

[hgm]

The first cycle of this 55+5 tournament is now completed. In this cycle the programs in principle played two games of Gothic Chess against each other. (Programs not licenced for Gothic Chess played from an opening that had Archbishop and Chancellor interchanged in stead.)

Code: Select all

                              Jo TJ Sm Sm TS Fa Bi Ar
 1. Joker80 n                 ## 11 11 10 11 11 11 11   93%  13.0 ( 86.0,  79.5)
 2. TJchess10x8               00 ## 0= 11 11 11 11 01   68%   9.5 ( 93.0,  51.3)
 3. Smirf 1.73f-X             00 1= ## 10 10 11 01 11   61%   8.5 ( 95.0,  47.8)
 4. Smirf Donation            01 00 01 ## 10 =0 10 11   46%   6.5 ( 99.0,  40.3)
 5. TSCP Gothic               00 00 01 01 ## 10 10 11   43%   6.0 (100.0,  30.5)
 6. Fairy-Max 4.8 v           00 00 00 =1 01 ## 11 10   39%   5.5 (101.0,  26.8)
 7. BigLion 15mar             00 00 10 01 01 00 ## 01   29%   4.0 (104.0,  24.0)
 8. ArcBishop80 1.00          00 10 00 00 00 01 10 ##   21%   3.0 (106.0,  19.0)

I just started cycle #2, in which all programs play each other twice at Bird Chess.

[hgm]

The second cycle (2 games of Bird Chess) of this tournament is now completed. Joker80 and Smirf 1.73g-X did not give each other an inch, this time. For the nr. 2 after the first cycle, TJchess10x8, this was a very bad cycle: it lost 4 points compared to both Joker80 and Smirf 1.73g-X. It did still better than all the others, though.

Current standing:

Code: Select all

                              Joke Smir TJch TSCP Smir Fair ArcB BigL
 1. Joker80 n                 #### 1101 1111 1111 1001 1111 1111 1111   89%  25.0 (348.0, 301.5)
 2. Smirf 1.73g-X             0010 #### 1=11 1010 1011 1111 1111 0111   73%  20.5 (366.0, 236.3)
 3. TJchess10x8               0000 0=00 #### 1111 1110 1110 0111 1111   62%  17.5 (378.0, 176.8)
 4. TSCP Gothic               0000 0101 0000 #### 0111 10=1 111= 10=1   48%  13.5 (394.0, 143.3)
 5. Smirf Donation            0110 0100 0001 1000 #### =011 1110 101=   46%  13.0 (396.0, 161.3)
 6. Fairy-Max 4.8 v           0000 0000 0001 01=0 =100 #### 100= 1111   34%   9.5 (410.0,  91.8)
 7. ArcBishop80 1.00          0000 0000 1000 000= 0001 011= #### 1001   25%   7.0 (420.0,  73.0)
 8. BigLion80 4apr            0000 1000 0000 01=0 010= 0000 0110 ####   21%   6.0 (424.0,  74.3)

We now continue with a cycle of Capablanca Chess. Since TSCP Gothic cannot play that, however, it will play all its games later, from a fantasy opening setup that can be reached from the Gothic opening position (which is hard-wired into TSCP-G).

[smrf]

Thank you Harm for making this event possible. Moreover now it has become obvious, how valuable it has been, that you wrote the Smirf-o-Glot adapter. Without SMIRF here the event would be a boring Joker apotheosis. Nevertheless SMIRF in its current design seems not yet to be able to compete with Joker.

All the best for your event and Joker, Reinhard.

[hgm]

You caught me there! This is indeed why I did it.

This event would not have been very interesting without Smirf participating. I was waiting for the new ChessV (which should be WB compatible), but it was delayed. Also for testing Joker I was badly in need of a tougher opponent.

Many engines, though, still throw away games against weaker opponents because they make unforced losing trades. I think this is the main reason why Joker80 seems to be so much ahead of the pack, now.

[hgm]

The 3rd cycle is now completed. Most games were Capablanca Chess in this cycle; only TSCP Gothic, which is incapable of playing this initial setup even through kludges, did play all its games with the opening array RABNQKNBCR. In the next cycle we will play from the (mirrored) Embassy-Chess setup; This will pose a similar problem, which is why TSCP-G will pre-play its games for this cycle with the opening array RNBAQKCBNR.

Code: Select all

                              Jok Smi TJc Smi TSC Fai Big Arc
 1. Joker80 n                 ### 110 111 100 111 111 111 111
                              ### 100 101 101 111 111 111 111   83%  35.0 (798.0, 619.5)

 2. Smirf 1.73g-X             001 ### 1=1 101 101 111 011 111
                              011 ### 100 111 000 101 111 111   70%  29.5 (831.0, 529.2)

 3. TJchess10x8               000 0=0 ### 111 111 111 111 011
                              010 011 ### 01= 11= 011 110 111   68%  28.5 (837.0, 483.5)

 4. Smirf Donation            011 010 000 ### 100 =01 101 111
                              010 000 10= ### 010 101 =11 011   49%  20.5 (885.0, 362.8)

 5. TSCP Gothic               000 010 000 011 ### 10= 10= 111
                              000 111 00= 101 ### 100 110 =11   48%  20.0 (888.0, 339.5)

 6. Fairy-Max 4.8 v           000 000 000 =10 01= ### 111 100
                              000 010 100 010 011 ### 110 =00   35%  14.5 (921.0, 246.0)

 7. BigLion80 4apr            000 100 000 010 01= 000 ### 011
                              000 000 001 =00 001 001 ### 0=1   25%  10.5 (945.0, 186.5)

 8. ArcBishop80 1.00          000 000 100 000 000 011 100 ###
                              000 000 000 100 =00 =11 1=0 ###   23%   9.5 (951.0, 150.5)

Despite losing 0-2 to the runner-up Smirf Bonus in this cycle, Joker80 was able to build out its lead by one point, due to Smirf losing 2-0 against both TJchess10x8 and TSCP Gothic. Both Joker80 and Smirf Bonus were struck by a horrible bug, which caused them to throw away a game when they already saw how to checkmate their opponent. Due to this, and its excellent own results, TJchess10x8 was able to gain back some of the ground it lost in the previous cycle, and now approached Smirf Bonus to a single point.

[Matthias Gemuh]

Code: Select all

                              Jok Smi TJc Smi TSC Fai Big Arc
 1. Joker80 n                 ### 110 111 100 111 111 111 111
                              ### 100 101 101 111 111 111 111   83%  35.0 (798.0, 619.5)

 2. Smirf 1.73g-X             001 ### 1=1 101 101 111 011 111
                              011 ### 100 111 000 101 111 111   70%  29.5 (831.0, 529.2)

 3. TJchess10x8               000 0=0 ### 111 111 111 111 011
                              010 011 ### 01= 11= 011 110 111   68%  28.5 (837.0, 483.5)

 4. Smirf Donation            011 010 000 ### 100 =01 101 111
                              010 000 10= ### 010 101 =11 011   49%  20.5 (885.0, 362.8)

 5. TSCP Gothic               000 010 000 011 ### 10= 10= 111
                              000 111 00= 101 ### 100 110 =11   48%  20.0 (888.0, 339.5)

 6. Fairy-Max 4.8 v           000 000 000 =10 01= ### 111 100
                              000 010 100 010 011 ### 110 =00   35%  14.5 (921.0, 246.0)

 7. BigLion80 4apr            000 100 000 010 01= 000 ### 011
                              000 000 001 =00 001 001 ### 0=1   25%  10.5 (945.0, 186.5)

 8. ArcBishop80 1.00          000 000 100 000 000 011 100 ###
                              000 000 000 100 =00 =11 1=0 ###   23%   9.5 (951.0, 150.5)

BigLion80 and ArcBishop80 are searching at 5%-15% speed compared to the other engines because of my slow home-made 10x8-Bitboard class (no Rybka node count obfuscation) and searching 5 plys
shallower. Yet they sometimes grab a point against the monsters .
Who can explain this phenomenon ? Can it be that BigLion80 and ArcBishop80 have some advantageous bugs ?

Matthias.

[Uri Blass]

Code: Select all

                              Jok Smi TJc Smi TSC Fai Big Arc
 1. Joker80 n                 ### 110 111 100 111 111 111 111
                              ### 100 101 101 111 111 111 111   83%  35.0 (798.0, 619.5)

 2. Smirf 1.73g-X             001 ### 1=1 101 101 111 011 111
                              011 ### 100 111 000 101 111 111   70%  29.5 (831.0, 529.2)

 3. TJchess10x8               000 0=0 ### 111 111 111 111 011
                              010 011 ### 01= 11= 011 110 111   68%  28.5 (837.0, 483.5)

 4. Smirf Donation            011 010 000 ### 100 =01 101 111
                              010 000 10= ### 010 101 =11 011   49%  20.5 (885.0, 362.8)

 5. TSCP Gothic               000 010 000 011 ### 10= 10= 111
                              000 111 00= 101 ### 100 110 =11   48%  20.0 (888.0, 339.5)

 6. Fairy-Max 4.8 v           000 000 000 =10 01= ### 111 100
                              000 010 100 010 011 ### 110 =00   35%  14.5 (921.0, 246.0)

 7. BigLion80 4apr            000 100 000 010 01= 000 ### 011
                              000 000 001 =00 001 001 ### 0=1   25%  10.5 (945.0, 186.5)

 8. ArcBishop80 1.00          000 000 100 000 000 011 100 ###
                              000 000 000 100 =00 =11 1=0 ###   23%   9.5 (951.0, 150.5)

BigLion80 and ArcBishop80 are searching at 5%-15% speed compared to the other engines because of my slow home-made 10x8-Bitboard class (no Rybka node count obfuscation) and searching 5 plys
shallower. Yet they sometimes grab a point against the monsters .
Who can explain this phenomenon ? Can it be that BigLion80 and ArcBishop80 have some advantageous bugs ?

Matthias.

being 7-20 times slower does not mean infinite advantage in rating and the weaker player can score points so I see nothing strange about it.

another point is that 10*8 programs are relatively weak because of lack of interest in this game.

I believe that the best 10*8 program of 2009 is going to be more than 300 elo better than Joker of this tournament.

If Vas decide to write 10*8 program for 2009 then my prediction is that the difference is going to be more than 500 elo.

Uri

[Matthias Gemuh]

another point is that 10*8 programs are relatively weak because of lack of interest in this game.

Uri

The weakness of current 10x8 engines is due to lack of time, not lack of interest. Beating so many engines at 5 plys less in so few games still remains
a mystery to me, especially because my engines are poor in speed, search, eval etc. I hope they shall also grab points against Movei80

Matthias.

[smrf]

Indeed SMIRF still has some serious bugs caused by its original ad hoc design. Some minor improvements are now to be found in its renewed version MS-173h at http://www.chessbox.de/Compu/schachsmirf_e.html, downloadable for supporters and donators.

Nevertheless SMIRF has to be rewritten, for which already a lot of improving ideas exist. But there still is only small interest and not that much private time left for me to go on with it.

Reinhard.

[hgm]

BigLion80 and ArcBishop80 are searching at 5%-15% speed compared to the other engines because of my slow home-made 10x8-Bitboard class (no Rybka node count obfuscation) and searching 5 plys
shallower. Yet they sometimes grab a point against the monsters .
Who can explain this phenomenon ? Can it be that BigLion80 and ArcBishop80 have some advantageous bugs ?

The factor you quote is not enough to explain 5 ply of search depth difference. I suspect that your engines count depth differnt from the others. Both my engines use LMR, which makes them report 2 ply more depth at the same search ime. But the 2 ply is only along the PV, and the side branches are on the average even more than 2 ply shallower, making it just as susceptible to tactical surprises as without the LMR. In fact for Joker the LMR hardly seemed to add any strength.

Furthermore, most engines actually search deeper than they realize, because a QS node that fails high through a capture is actually equivaent to a 1-ply search. Joker's search is fully depth-aware, and it reports this extra depth, where other engines, with the same search tree, might report one ply less.

Despite all that, I am surprised about the low depth yuor engines report. It is almost like they are thousands of times slower. But it might be that you have more extensions than the other engines, so that your search isn't as shallow as the reported depth suggests.

Furthermore, your engines do seem to have better evaluation than most of the other participants: TSCP's evaluation is truly appalling, and most other engines seem to have issues with piece values (discussed extensively in the other thread). So sometimes they use their larger depth only to force the opponent to capture a piece worth of material. And then even the 4 extra ply can't save them anymore, at such long time controls.

I certainly hope that next year's winner will be 300 Elo stronger. If only because I plan to improve Joker that much!