search

jnlt3 / blackmarlin

WIP Chess Engine
GNU General Public License v3.0
92 stars 7 forks source link

Average eval and beta for RFP #242

Closed jnlt3 closed 6 months ago

jnlt3 commented 6 months ago

Passed STC:

Elo   | 4.52 +- 3.89 (95%)
SPRT  | 8.0+0.08s Threads=1 Hash=8MB
LLR   | 2.95 (-2.94, 2.94) [0.00, 4.00]
Games | N: 15082 W: 3821 L: 3625 D: 7636
Penta | [147, 1756, 3564, 1902, 172]

Passed LTC:

Elo   | 3.06 +- 3.09 (95%)
SPRT  | 40.0+0.40s Threads=1 Hash=64MB
LLR   | 2.96 (-2.94, 2.94) [0.00, 4.00]
Games | N: 22264 W: 5209 L: 5013 D: 12042
Penta | [66, 2499, 5829, 2649, 89]
tissatussa commented 6 months ago

i see the concerning function now returns (eval + beta) / 2 , not just eval .. i wonder : how does this influence the playing strength and style ? I feel changing that little code line will have a huge impact ..

jnlt3 commented 5 months ago

i see the concerning function now returns (eval + beta) / 2 , not just eval .. i wonder : how does this influence the playing strength and style ? I feel changing that little code line will have a huge impact ..

No way to tell how it influences style as it is not quantifiable. It does increase strength as shown by the SPRTs though. It is certainly an interesting way to fail high but it is still logically sound as eval is guaranteed to be greater than or equal beta and thus the mean value of the two is guaranteed to be greater than or equal to beta too.