Author: Chris Bolt Rating CEDR=2493
Eubos 4.3 what's new?
Main effort was to optimise board updates when playing moves and to fix a bug in multithreading which meant that any extra thread from the main used the wrong hash after a castling move occurred in the game, thereby ruining the effect of lazy SMP. I spent a lot of effort trying to not update the iterative neural net evaluation when a move resulted in an insufficient draw, but this had negligible, perhaps detrimental effect, so that work was reverted.
Aslo, the neural net was retrained using the same training data as per v4.2, but with a smaller batch size, resulting in a small strength improvement by reducing over-training. I am still experimenting with the learning rate scheduler to optimise performance for this training set.
Eubos 3.9 JA vs other engines:
| Pioneer 0.3.2 JA | 4/4 | +4 | 4 Games |
| Mufasa 0.2.1 JA | 3.5/4 | +3 | 4 Games |
| Integral 7.0.0 | 0/4 | -4 | 4 Games |
| HypnoS ++ 1.02 | 0/4 | -4 | 4 Games |
| GOOB 1.8.9 JA | 0/4 | -4 | 4 Games |
| Peacekeeper 3.01 JA | 0/4 | -4 | 4 Games |
| Odonata 1.0.0 JA | 0/4 | -4 | 4 Games |
| Obsidian 15.03 SE | 0/4 | -4 | 4 Games |
| Killfish PB 090225 | 0/4 | -4 | 4 Games |
| Midnight 9 JA | 0/4 | -4 | 4 Games |
| Fatalii 0.9.0 | 0/4 | -4 | 4 Games |
| Beef 0.3.6 JA | 0/3 | -3 | 3 Games |
| Bitbit 1.3 | 2/2 | +2 | 2 Games |
| Rustic Alpha 3.05 JA | 2/2 | +2 | 2 Games |
| Reggz 0.6.0 | 2/2 | +2 | 2 Games |
| Deep Blunder 1.2.0 JA | 2/2 | +2 | 2 Games |
| Jester 0.85 JA | 2/2 | +2 | 2 Games |
| Sloth 2.0 JA | 1.5/2 | +1 | 2 Games |
| 100 Elo Chess Engine | 1.5/2 | +1 | 2 Games |
| Fatalii 0.9.0 JA | 1/2 | +0 | 2 Games |
| Stro4k 3.0 JA | 1/2 | +0 | 2 Games |
Jim Ablett compiles:

Comments
Post a Comment