Toga II 4.0 release

[jd1]

Dear all,

Happy New Year! It is my pleasure, after a break of several years from computer chess, to release Toga II 4.0.

Strength

Self-play results:
+80 elo at 4' + 0.04' on 1 thread after 4000 games
+85 elo at 10' + 0.1' on 4 threads after 2000 games

Nice job.

Some months ago, just for the fun of it, I looked into the eval of Toga II 3.0 and found some improvements. By head I remember -

#1. Instead of fixed values for isolated and backward pawns use a PST.

#2. Remove the own pawn support bonus for knight outposts.

#3. Bishop pair via table[number of pawns] instead of fixed values

#4. Bishop mobility, instead of only looking at the 2 forward directions do all 4 directions after all

#5. Also throwing out the 2nd Lazy Eval Cutoff gave elo.

Wasn't able to improve on King Safety, hated that.

Thank you so much Ed!

I will try #1 to #3, #1 and #3 in particular seem like they should be worth significant Elo.

Toga II 4.0 already incorporates #4.

As for #5, could I please ask how many games you tested it with and at what time control? I twice ran tests to try to cut it out ~(3000, then 8000 games at 3' + 0.03') but never managed to get even a plus score by removing it.

[De Noose Daniel]

I made a little test in 1' + 0,1'/move with some of the most recent Toga II versions :

Code: Select all

Games Completed = 1200 of 1200 (Avg game length = 111.667 sec)
Settings = Gauntlet/128MB/60000ms+100ms/M 1000cp for 12 moves, D 150 moves/EPD:Hyatt_4000_openings.epd(4000)
Time = 54012 sec elapsed, 0 sec remaining
 1.  Toga II 4.01             	807.5/1200	602-187-411  	(L: m=73 t=0 i=0 a=114)	(D: r=270 i=63 f=54 s=1 a=23)	(tpm=792.7 d=16.62 nps=715642)
 2.  Toga II Braesch          	88.0/200	44-68-88  	(L: m=26 t=0 i=0 a=42)	(D: r=56 i=14 f=11 s=0 a=7)	(tpm=825.0 d=15.26 nps=693030)
 3.  Toga II Hair             	53.0/200	37-131-32  	(L: m=52 t=0 i=0 a=79)	(D: r=17 i=4 f=7 s=0 a=4)	(tpm=871.7 d=14.93 nps=689407)
 4.  Toga II Michniewski      	58.0/200	29-113-58  	(L: m=52 t=0 i=0 a=61)	(D: r=37 i=9 f=10 s=0 a=2)	(tpm=860.6 d=15.34 nps=685228)
 5.  Toga Returns 1.0         	52.5/200	18-113-69  	(L: m=39 t=0 i=0 a=74)	(D: r=40 i=13 f=12 s=0 a=4)	(tpm=855.6 d=13.19 nps=808766)
 6.  Toga II 2.02             	60.5/200	21-100-79  	(L: m=30 t=0 i=0 a=70)	(D: r=64 i=8 f=4 s=1 a=2)	(tpm=909.7 d=13.50 nps=818818)
 7.  Toga II 3.0              	80.5/200	38-77-85  	(L: m=28 t=0 i=0 a=49)	(D: r=56 i=15 f=10 s=0 a=4)	(tpm=835.5 d=15.04 nps=750982)

Toga II 4.01 is clearly better than all them !

As Jerry said, it seems nearly 70 elo better than Toga II 3.0 in this test.

I'm a bit astonished about the Toga Braesch who plays better than Toga II 3.0.

Maybe it'll be interesting to test the Lucas Braesch PST in the new Toga II 4.01

[Rebel]

Dear all,

Happy New Year! It is my pleasure, after a break of several years from computer chess, to release Toga II 4.0.

Strength

Self-play results:
+80 elo at 4' + 0.04' on 1 thread after 4000 games
+85 elo at 10' + 0.1' on 4 threads after 2000 games

Nice job.

Some months ago, just for the fun of it, I looked into the eval of Toga II 3.0 and found some improvements. By head I remember -

#1. Instead of fixed values for isolated and backward pawns use a PST.

#2. Remove the own pawn support bonus for knight outposts.

#3. Bishop pair via table[number of pawns] instead of fixed values

#4. Bishop mobility, instead of only looking at the 2 forward directions do all 4 directions after all

#5. Also throwing out the 2nd Lazy Eval Cutoff gave elo.

Wasn't able to improve on King Safety, hated that.

Thank you so much Ed!

I will try #1 to #3, #1 and #3 in particular seem like they should be worth significant Elo.

Toga II 4.0 already incorporates #4.

As for #5, could I please ask how many games you tested it with and at what time control? I twice ran tests to try to cut it out ~(3000, then 8000 games at 3' + 0.03') but never managed to get even a plus score by removing it.

Code: Select all

#define LAZY_EVAL_2		OFF			// 51.0% | los=95.3% 

is all I have. I usually play 4800 games at 40/60 in case of search changes, so 1.5 sec per move average.

[jd1]

Dear all,

Happy New Year! It is my pleasure, after a break of several years from computer chess, to release Toga II 4.0.

Strength

Self-play results:
+80 elo at 4' + 0.04' on 1 thread after 4000 games
+85 elo at 10' + 0.1' on 4 threads after 2000 games

Nice job.

Some months ago, just for the fun of it, I looked into the eval of Toga II 3.0 and found some improvements. By head I remember -

#1. Instead of fixed values for isolated and backward pawns use a PST.

#2. Remove the own pawn support bonus for knight outposts.

#3. Bishop pair via table[number of pawns] instead of fixed values

#4. Bishop mobility, instead of only looking at the 2 forward directions do all 4 directions after all

#5. Also throwing out the 2nd Lazy Eval Cutoff gave elo.

Wasn't able to improve on King Safety, hated that.

Thank you so much Ed!

I will try #1 to #3, #1 and #3 in particular seem like they should be worth significant Elo.

Toga II 4.0 already incorporates #4.

As for #5, could I please ask how many games you tested it with and at what time control? I twice ran tests to try to cut it out ~(3000, then 8000 games at 3' + 0.03') but never managed to get even a plus score by removing it.

Code: Select all

#define LAZY_EVAL_2		OFF			// 51.0% | los=95.3% 

is all I have. I usually play 4800 games at 40/60 in case of search changes, so 1.5 sec per move average.

Thanks Ed. That's a far longer TC than what I used, that must be the reason. Lazy eval doesn't tend to scale well in Toga II at least.

[jd1]

I made a little test in 1' + 0,1'/move with some of the most recent Toga II versions :

Code: Select all

Games Completed = 1200 of 1200 (Avg game length = 111.667 sec)
Settings = Gauntlet/128MB/60000ms+100ms/M 1000cp for 12 moves, D 150 moves/EPD:Hyatt_4000_openings.epd(4000)
Time = 54012 sec elapsed, 0 sec remaining
 1.  Toga II 4.01             	807.5/1200	602-187-411  	(L: m=73 t=0 i=0 a=114)	(D: r=270 i=63 f=54 s=1 a=23)	(tpm=792.7 d=16.62 nps=715642)
 2.  Toga II Braesch          	88.0/200	44-68-88  	(L: m=26 t=0 i=0 a=42)	(D: r=56 i=14 f=11 s=0 a=7)	(tpm=825.0 d=15.26 nps=693030)
 3.  Toga II Hair             	53.0/200	37-131-32  	(L: m=52 t=0 i=0 a=79)	(D: r=17 i=4 f=7 s=0 a=4)	(tpm=871.7 d=14.93 nps=689407)
 4.  Toga II Michniewski      	58.0/200	29-113-58  	(L: m=52 t=0 i=0 a=61)	(D: r=37 i=9 f=10 s=0 a=2)	(tpm=860.6 d=15.34 nps=685228)
 5.  Toga Returns 1.0         	52.5/200	18-113-69  	(L: m=39 t=0 i=0 a=74)	(D: r=40 i=13 f=12 s=0 a=4)	(tpm=855.6 d=13.19 nps=808766)
 6.  Toga II 2.02             	60.5/200	21-100-79  	(L: m=30 t=0 i=0 a=70)	(D: r=64 i=8 f=4 s=1 a=2)	(tpm=909.7 d=13.50 nps=818818)
 7.  Toga II 3.0              	80.5/200	38-77-85  	(L: m=28 t=0 i=0 a=49)	(D: r=56 i=15 f=10 s=0 a=4)	(tpm=835.5 d=15.04 nps=750982)

Toga II 4.01 is clearly better than all them !

As Jerry said, it seems nearly 70 elo better than Toga II 3.0 in this test.

I'm a bit astonished about the Toga Braesch who plays better than Toga II 3.0.

Maybe it'll be interesting to test the Lucas Braesch PST in the new Toga II 4.01

Thanks a lot for the test!

Yes, I'll try Lucas' values in the latest Toga to see if there's any improvement.

Note that Toga II Braesch, Hair and Michniewski all incorporate the improvements from Toga II 3.0. So it's not too surprising that one managed to do better than the default PST in Toga II 3.0 in 200 games.

[De Noose Daniel]

Thanks to take a look at this. Very nice.

(sorry for my bad english)


Here is a faster test (10" + 50ms) showing again a plus with Braesch PST but less efficient.

Code: Select all

Games Completed = 8000 of 8000 (Avg game length = 22.133 sec)
Settings = Gauntlet/128MB/10000ms+50ms/M 1000cp for 12 moves, D 150 moves/EPD:Hyatt_4000_openings.epd(4000)
Time = 79137 sec elapsed, 0 sec remaining
 1.  Toga II 4.01             	4840.0/8000	3350-1670-2980  	(L: m=980 t=0 i=0 a=690)	(D: r=2050 i=486 f=296 s=22 a=126)	(tpm=164.2 d=12.58 nps=840888)
 2.  Toga II Braesch          	1610.5/4000	894-1673-1433  	(L: m=809 t=0 i=0 a=864)	(D: r=939 i=258 f=149 s=12 a=75)	(tpm=168.9 d=11.56 nps=785068)
 3.  Toga II 3.0              	1549.5/4000	776-1677-1547  	(L: m=807 t=0 i=0 a=870)	(D: r=1111 i=228 f=147 s=10 a=51)	(tpm=172.3 d=11.59 nps=895526)

This test seems to show an equal number of losses but clearly more wins than draws.

[tomitank]

Maybe it'll be interesting to test the Lucas Braesch PST in the new Toga II 4.01

Where can I download this version? (Toga - Lucas Braesch)

[jd1]

I made a little test in 1' + 0,1'/move with some of the most recent Toga II versions :

Code: Select all

Games Completed = 1200 of 1200 (Avg game length = 111.667 sec)
Settings = Gauntlet/128MB/60000ms+100ms/M 1000cp for 12 moves, D 150 moves/EPD:Hyatt_4000_openings.epd(4000)
Time = 54012 sec elapsed, 0 sec remaining
 1.  Toga II 4.01             	807.5/1200	602-187-411  	(L: m=73 t=0 i=0 a=114)	(D: r=270 i=63 f=54 s=1 a=23)	(tpm=792.7 d=16.62 nps=715642)
 2.  Toga II Braesch          	88.0/200	44-68-88  	(L: m=26 t=0 i=0 a=42)	(D: r=56 i=14 f=11 s=0 a=7)	(tpm=825.0 d=15.26 nps=693030)
 3.  Toga II Hair             	53.0/200	37-131-32  	(L: m=52 t=0 i=0 a=79)	(D: r=17 i=4 f=7 s=0 a=4)	(tpm=871.7 d=14.93 nps=689407)
 4.  Toga II Michniewski      	58.0/200	29-113-58  	(L: m=52 t=0 i=0 a=61)	(D: r=37 i=9 f=10 s=0 a=2)	(tpm=860.6 d=15.34 nps=685228)
 5.  Toga Returns 1.0         	52.5/200	18-113-69  	(L: m=39 t=0 i=0 a=74)	(D: r=40 i=13 f=12 s=0 a=4)	(tpm=855.6 d=13.19 nps=808766)
 6.  Toga II 2.02             	60.5/200	21-100-79  	(L: m=30 t=0 i=0 a=70)	(D: r=64 i=8 f=4 s=1 a=2)	(tpm=909.7 d=13.50 nps=818818)
 7.  Toga II 3.0              	80.5/200	38-77-85  	(L: m=28 t=0 i=0 a=49)	(D: r=56 i=15 f=10 s=0 a=4)	(tpm=835.5 d=15.04 nps=750982)

Toga II 4.01 is clearly better than all them !

As Jerry said, it seems nearly 70 elo better than Toga II 3.0 in this test.

I'm a bit astonished about the Toga Braesch who plays better than Toga II 3.0.

Maybe it'll be interesting to test the Lucas Braesch PST in the new Toga II 4.01

So I looked at the readme and the sources ... Toga II Braesch is based on Toga II 280513 which is an old beta of mine that is 20-30 elo stronger than version 3.0. That's why it consistently performs better than Toga II 3.0, it's not really (or mainly) the piece-square tables.

[jd1]

Dear all,

Happy New Year! It is my pleasure, after a break of several years from computer chess, to release Toga II 4.0.

Strength

Self-play results:
+80 elo at 4' + 0.04' on 1 thread after 4000 games
+85 elo at 10' + 0.1' on 4 threads after 2000 games

Nice job.

Some months ago, just for the fun of it, I looked into the eval of Toga II 3.0 and found some improvements. By head I remember -

#1. Instead of fixed values for isolated and backward pawns use a PST.

#2. Remove the own pawn support bonus for knight outposts.

#3. Bishop pair via table[number of pawns] instead of fixed values

#4. Bishop mobility, instead of only looking at the 2 forward directions do all 4 directions after all

#5. Also throwing out the 2nd Lazy Eval Cutoff gave elo.

Wasn't able to improve on King Safety, hated that.

I've now tested all of them in Toga II 4.0:

#1. Was an improvement

#2. Lost elo every time I tried. However Toga II 4.0 mobility eval is quite different to 3.0, so there was no guarantee that it would work.

#3. Was a slight improvement

#4. N/A, already in

#5. Couldn't get this to work either, 48.4% after 6000 games. Toga II 4.0 has an even slower mobility function than 3.0 so perhaps the second cutoff becomes useful.

[supersharp77]

Dear all,

Happy New Year! It is my pleasure, after a break of several years from computer chess, to release Toga II 4.0.

Strength

Self-play results:
+80 elo at 4' + 0.04' on 1 thread after 4000 games
+85 elo at 10' + 0.1' on 4 threads after 2000 games

Nice job.

Some months ago, just for the fun of it, I looked into the eval of Toga II 3.0 and found some improvements. By head I remember -

#1. Instead of fixed values for isolated and backward pawns use a PST.

#2. Remove the own pawn support bonus for knight outposts.

#3. Bishop pair via table[number of pawns] instead of fixed values

#4. Bishop mobility, instead of only looking at the 2 forward directions do all 4 directions after all

#5. Also throwing out the 2nd Lazy Eval Cutoff gave elo.

Wasn't able to improve on King Safety, hated that.

Thank you so much Ed!

I will try #1 to #3, #1 and #3 in particular seem like they should be worth significant Elo.

Toga II 4.0 already incorporates #4.

As for #5, could I please ask how many games you tested it with and at what time control? I twice ran tests to try to cut it out ~(3000, then 8000 games at 3' + 0.03') but never managed to get even a plus score by removing it.

Code: Select all

#define LAZY_EVAL_2		OFF			// 51.0% | los=95.3% 

is all I have. I usually play 4800 games at 40/60 in case of search changes, so 1.5 sec per move average.

Thanks Ed. That's a far longer TC than what I used, that must be the reason. Lazy eval doesn't tend to scale well in Toga II at least.

At 18-19 ply depth (Toga II v 4.0) Toga is now quite dangerous opponent for almost any chess engine in the world...Undefeated in my current RR tourney after 7-8 rounds....Look Out!!