The Baron 2.23 without, and with LMR

[Mike S.]

4m+1s, AthlonXP@1439 MHz
128 MB hash, ponder off
4 short predefined opening variations
4-piece tbs. (bitbases for Bright)
Fritz 8-GUI, Windows 98SE

1. without LMR (lmrmoves 0); that's the default

Code: Select all

The Baron 2.23   - Bright-0.3a        2.5 - 5.5    +1/-4/=3    31.25%
The Baron 2.23   - Fruit 051103       0.5 - 7.5    +0/-7/=1    6.25%
The Baron 2.23   - Nimzo 8            2.5 - 5.5    +2/-5/=1    31.25%
The Baron 2.23   - Hiarcs 9           2.5 - 5.5    +1/-4/=3    31.25%

total 8,0/32 = 25,0%

(all learn files were deleted between the two tests)

2. with LMR (lmrmoves 3); other lmr settings remained unchanged

Code: Select all

The Baron 2.23   - Bright-0.3a        2.5 - 5.5    +1/-4/=3    31.25%
The Baron 2.23   - Fruit 051103       1.5 - 6.5    +1/-6/=1    18.75%
The Baron 2.23   - Nimzo 8            4.5 - 3.5    +3/-2/=3    56.25%
The Baron 2.23   - Hiarcs 9           2.5 - 5.5    +2/-5/=1    31.25%

total 11,0/32 = 34,4%

The effect of LMR is more visible if depths under equal conditions are compared. Here are some examples from these games, always without LMR at first, and then with LMR:

[White "The Baron 2.23"]
[Black "Bright-0.3a"]

1. d4 Nf6 2. c4 e6 3. Nf3 b6 4. g3 Bb4+ {0.03/11 9} 5. Bd2 {-0.08/10 13} Bxd2+
{(Sc6) 0.09/13 15} 6. Qxd2 {(Sbxd2) 0.16/10 10} Ba6 {(Sc6) 0.21/13 29} 7. Nc3 {
(b3) 0.11/9 6} Bxc4 {(Sc6) -0.41/13 9} 8. Bg2 {-0.13/10 18}

[White "The Baron 2.23"]
[Black "Bright-0.3a"]

1. d4 Nf6 2. c4 e6 3. Nf3 b6 4. g3 Bb4+ {0.03/11 9} 5. Bd2 {0.04/11 16} Bxd2+ {
(Sc6) 0.09/13 15} 6. Nbxd2 {-0.04/12 10} Bb7 {(Sc6) 0.04/14 12} 7. Bg2 {
0.04/11 10} O-O {(Sc6) 0.04/13 8} 8. O-O {0.05/11 8}

[White "The Baron 2.23"]
[Black "Fruit 051103"]

1. d4 Nf6 2. c4 e6 3. Nf3 b6 4. g3 Bb4+ {-0.05/12 10} 5. Bd2 {0.07/11 12} Bxd2+
{(Sc6) 0.01/13 13} 6. Qxd2 {(Sbxd2) 0.11/10 11} Ba6 {-0.13/12 7} 7. Nc3 {
(b3) 0.09/9 9} Bxc4 {(0-0) -0.30/13 8} 8. Ne5 {(Lg2) -0.17/10 13}

[White "The Baron 2.23"]
[Black "Fruit 051103"]

1. d4 Nf6 2. c4 e6 3. Nf3 b6 4. g3 Bb4+ {-0.05/12 10} 5. Bd2 {-0.07/12 15}
Bxd2+ {(Sc6) 0.00/13 10} 6. Qxd2 {(Sbxd2) 0.13/11 9} Ba6 {(Sc6) -0.06/12 8} 7.
Nc3 {(b3) -0.06/10 5} Bxc4 {-0.34/13 9} 8. Bg2 {-0.16/10 5}

[White "The Baron 2.23"]
[Black "Nimzo 8"]

1. e4 c5 2. Nf3 d6 3. d4 cxd4 4. Nxd4 Nf6 {0.13/10 11} 5. Nc3 {0.97/9 20} e5 {
(Sc6) 0.11/10 11} 6. Bb5+ {1.11/9 9} Bd7 {(Sbd7) 0.44/11 8} 7. Bxd7+ {1.28/9 5}
Qxd7 {0.23/10 7} 8. Nf5 {(Sf3) 1.07/9 7}

[White "The Baron 2.23"]
[Black "Nimzo 8"]

1. e4 c5 2. Nf3 d6 3. d4 cxd4 4. Nxd4 Nf6 {0.29/10 13} 5. Nc3 {0.90/10 16} e5 {
(Sc6) 0.11/10 11} 6. Bb5+ {1.18/10 7} Bd7 {(Sbd7) 0.44/11 9} 7. Bxd7+ {
1.15/11 11} Qxd7 {0.23/10 7} 8. Nf5 {(Sf3) 0.99/11 12}

Therefore, I think it seems recommendable to activate LMR in Baron 2.23.

[Dr.Wael Deeb]

Thanks Mike for this quick research
I also think that it would be nice if Richard can elaborate a little regarding this issue....

[pijl]

Thanks for testing. Although this quick test seems to indicate that LMR might work for the Baron I urge everybody to be careful with it.
If you look at the search depths with/without LMR you see that it gains only 1 reported ply, but it might reduce more than that in critical tactical lines.
But perhaps it is not a big problem as the Baron is not that strong in tactics anyway.

Other tests you might try is to vary the lmrmargin parameter. lmrmargin 0 should give an even bigger depth gain, but will also be less careful in the choice of reduced lines.
Richard.

[Harald Johnsen]

Aren't the positions against Fruit & Bright the same ?
Why a different score when just the opponent is different ?

HJ.

[pijl]

Aren't the positions against Fruit & Bright the same ?
Why a different score when just the opponent is different ?

HJ.

Most likely due to position learning. As I understood the learning files were only cleared once, between the LMR and non-LMR runs.
Richard.

[Mike S.]

As I understood the learning files were only cleared once, between the LMR and non-LMR runs.

Yes. I noticed this difference too... The depths were different too, in these cases. LMR had the effect of +1 in addition to both of it.

I can add a fourth pair of examples, with the variation from Baron-Nimzo, which was for the first time played against Bright:

[White "The Baron 2.23"]
[Black "Bright-0.3a"]

1. e4 c5 2. Nf3 d6 3. d4 cxd4 4. Nxd4 Nf6 {0.56/12 13} 5. Nc3 {1.17/9 15} Nc6 {
(e5) 0.45/12 14} 6. Bb5 {(Sxc6) 1.13/9 12} Bd7 {0.39/12 11} 7. O-O {
(Le3) 0.92/8 6} g6 {(e6) 0.24/12 10} 8. Be3 {(Sxc6) 0.98/8 5}

[White "The Baron 2.23"]
[Black "Bright-0.3a"]

1. e4 c5 2. Nf3 d6 3. d4 cxd4 4. Nxd4 Nf6 {0.56/12 13} 5. Nc3 {1.28/10 11} Nc6
{(e5) 0.45/12 14} 6. Bb5 {(Sxc6) 1.05/10 11} Bd7 {0.39/12 11} 7. O-O {
(Le3) 0.90/9 6} g6 {(e6) 0.24/12 10} 8. Be3 {(Sxc6) 1.01/9 8}

Again, from the Nimzo games for comparison:

[White "The Baron 2.23"]
[Black "Nimzo 8"]

1. e4 c5 2. Nf3 d6 3. d4 cxd4 4. Nxd4 Nf6 {0.13/10 11} 5. Nc3 {0.97/9 20} e5 {
(Sc6) 0.11/10 11} 6. Bb5+ {1.11/9 9} Bd7 {(Sbd7) 0.44/11 8} 7. Bxd7+ {1.28/9 5}
Qxd7 {0.23/10 7} 8. Nf5 {(Sf3) 1.07/9 7}

[White "The Baron 2.23"]
[Black "Nimzo 8"]

1. e4 c5 2. Nf3 d6 3. d4 cxd4 4. Nxd4 Nf6 {0.29/10 13} 5. Nc3 {0.90/10 16} e5 {
(Sc6) 0.11/10 11} 6. Bb5+ {1.18/10 7} Bd7 {(Sbd7) 0.44/11 9} 7. Bxd7+ {
1.15/11 11} Qxd7 {0.23/10 7} 8. Nf5 {(Sf3) 0.99/11 12}

Here, we find eval differences again, for the no LMR and LMR pairs each, but not depth difference between them for 5.Nc3.

We also find a difference for Nimzo 8's evaluation after 4...Nf6. I did delete Nimzo's learning file too, between the test runs. It looks like Nimzo applies a small randomness for the time consumption, and changed it's eval at depth 10 between the 11th and 13th second.

The times are different for the Baron's moves too, so this could also be part of the explanantion (I don't know if Baron uses some randomness for the time consumption).

[pijl]

I don't know if Baron uses some randomness for the time consumption.

No randomness. It's due to the learning after the first game. It will use the scores from the previous games the next time it encounters the position in search, like a sort of permanent hashtable.
Richard.

[bob]

also the general consensus has been that the "history information" (I assume you are doing that) is not particularly effective, and is almost random noise thrown in to the equation.

There are other interesting aspects of "LMR" that can be tried as well, besides what is being done today by most everyone.

[Mike S.]

Thanks for your participation. - I am ready to repeat this test run with any other suggested settings for The Baron 2.23, if desired.

[Martin Thoresen]

Richard,

Pardon for asking another question in this thread but I simply cannot get The Baron 2.2.3 to use it's own opening book (the old one, same as 1.8.1 uses).

I have configured the two engines similarly but no matter what I do I can't get 2.2.3 to use it's own book nor can I activate the learning function, while 1.8.1 uses the book and learning without problems.

I've of course set the appropriate parameters in the .ini file.

Best Regards,
Martin