xr_a_y wrote: ↑Sun Jul 22, 2018 4:33 pm
So I decided I shall take some time to read again old, basic, stuff in the code... And of course, I've just discovered something bad, very bad. In QSearch, move generation was performed too soon, before standard cut-offs... Having fixed this, I ran a match versus fairy-max : the speed boost is quite interesting of course : more than 30% in knps ; but the elo gain seems crazy : now around >80% winrate (>250elo) versus fairy-max.
Rest assured, you are not the only one making stupid errors.
Recently I started tuning my old engine with Texels tuning method, and noticed that the parameters for contact checks didn't want to tune at all. Five years ago I changed something in the order in which the evaluation is done and that effectively broke contact checks. I overlooked this for 5 years, changing evaluation order should not affect things, this was pure sloppy programming from my side.
Texel tuning gave me ~72 additional Elo compared to my old tuning method which uses the 1500 STS positions, I didn't check the improvement with the repaired contact checks yet.
As I already knew, contact checks are real Elo makers. After the fix I ran 1000 games against the unfixed version and the Elo went up by 30 a 35 points. The improvement compared to the old situation is almost 100 Elo. The fun part is that I can check the performance against other engines this weekend during the HGM tourney.
Weini is now >90% win rate against fairy-max, thanks to search tuning (much more aggressive LMR, LMP and use of SEE in ordering). So it should be around 2300 elo now...
I'll test against Dorpsgek, Isa and Horizon now.
Still have to configure some other pruning and make use of SEE in search.
Cygwin issue shall be gone, at least it work for me.
