BikJump V3.1 is now available for download

[abik]

BikJump V3.1 is now available for download.

The following changes were made:

• improved evaluation function
• added castling/en-passant rights to zobrist key
• reverse and forward futility pruning
• scaled depth LMR, added LMP
• use move history and counter move in move ordering
• enhanced transposition table replacement policy
• tuned all parameters using hill climbing technique

Preliminary self-testing indicates that this version V3.1 is perhaps close to 200 ELO stronger than the recently releases V3 adnd about 400 ELO stronger than the older V2.0.1. But this is self-testing, so let's wait for some real results.

Against the older bikjump V3:

Code: Select all

--------------------------------------------------
Match    : ./bikjump300 vs. ./bikjump
Setting  : 400 games with 100 ms / move
--------------------------------------------------
Result : eng1 wins=71 eng2 wins=273 draws=56
Score  : eng2 301.0/400 = 75.25%
ELO    : eng1 -> eng2 = +193.17

Against the older bikjump V2.0.1:

Code: Select all

--------------------------------------------------
Match    : ./bikjump201 vs. ./bikjump
Setting  : 400 games with 100 ms / move
--------------------------------------------------
Result : eng1 wins=11 eng2 wins=345 draws=44
Score  : eng2 367.0/400 = 91.75%
ELO    : eng1 -> eng2 = +418.46

[Bart Weststrate]

tempo 60/20+1 Bikjump 3.1:

Score of SM_2182f vs Bikjump31: 240 - 181 - 79 [0.559] 500
... SM_2182f playing White: 134 - 79 - 37 [0.610] 250
... SM_2182f playing Black: 106 - 102 - 42 [0.508] 250
... White vs Black: 236 - 185 - 79 [0.551] 500
Elo difference: 41.2 +/- 28.1, LOS: 99.8 %, DrawRatio: 15.8 %

tempo 60/20+1 Bikjump 3.0:

Score of SM_2182b vs Bikjump: 320 - 128 - 52 [0.692] 500
... SM_2182b playing White: 162 - 68 - 20 [0.688] 250
... SM_2182b playing Black: 158 - 60 - 32 [0.696] 250
... White vs Black: 222 - 226 - 52 [0.496] 500
Elo difference: 140.6 +/- 31.0, LOS: 100.0 %, DrawRatio: 10.4 %

So +- 100 elo against the beancounter (which has a small bugfix since 2182b)

Match 60/60+1 is running.

added castling/en-passant rights to zobrist key

The bold is very important Aart. Especially if you search the hashmove before generating moves.

[abik]

So +- 100 elo against the beancounter (which has a small bugfix since 2182b)

Nice, although a bit lower than the self-tests indicated.
I appreciate your rapid response time with testing a lot Bart!

[Bart Weststrate]

60/60+1 Bikjump 3.1

Score of SM_2182f vs Bikjump31: 249 - 152 - 99 [0.597] 500
... SM_2182f playing White: 135 - 75 - 40 [0.620] 250
... SM_2182f playing Black: 114 - 77 - 59 [0.574] 250
... White vs Black: 212 - 189 - 99 [0.523] 500
Elo difference: 68.3 +/- 27.7, LOS: 100.0 %, DrawRatio: 19.8 %

Same against Bikjump 3.0:

Score of SM_2182b vs Bikjump: 343 - 91 - 66 [0.752] 500
... SM_2182b playing White: 181 - 41 - 28 [0.780] 250
... SM_2182b playing Black: 162 - 50 - 38 [0.724] 250
... White vs Black: 231 - 203 - 66 [0.528] 500
Elo difference: 192.7 +/- 32.1, LOS: 100.0 %, DrawRatio: 13.2 %

So +- 124 elo gained.

[abik]

60/60+1 Bikjump 3.1

Score of SM_2182f vs Bikjump31: 249 - 152 - 99 [0.597] 500
... SM_2182f playing White: 135 - 75 - 40 [0.620] 250
... SM_2182f playing Black: 114 - 77 - 59 [0.574] 250
... White vs Black: 212 - 189 - 99 [0.523] 500
Elo difference: 68.3 +/- 27.7, LOS: 100.0 %, DrawRatio: 19.8 %

Thanks Bart. So it looks I will have to find at least 70 ELO somewhere to go on par with SM_2182f.
I have some ideas left, but I am also reaching the diminishing returns stage for BikJump....

[abik]

I just posted a new Android engines package on Google Play that includes BikJump v2.5 (older version) and BikJump 3.1 (latest version). The binaries should import directly into Chess for Android on any armv7, arm64, or x86_64 Android device.

[abik]

A quick sanity tournament with 5s per move resulted in 10-0 in favor of BikJump v3.1!