I have yet to come up with the perfect implementation of the futility pruning in Mediocre. I ran a 300 game match between four different versions and the only conclusion was it is worth including captures in the futility, and I am already doing that. Using a limited material evaluation versus a full evaluation seems to be pretty much equal in strength.
There are a few more things to try out however, like not pruning pawns advancing to the seventh rank or similar, since they will most likely increase the score quite a bit.
There are also some adjustments to the evaluation that needs to be done. In v0.3 I had only a small bonus for rooks on the seventh rank since I did not check if there were actually pawns there (or opponent king on eight rank). Now in v0.31 I added that check so a rook on seventh should be valued higher than before.
The biggest flaw I found was with the x-ray attacks. A queen running diagonally into an own rook kept getting attack points behind it, this is obviously a bug and will be fixed.
Once I am happy with these things I will give candidate pawns another shot. I will most likely have to run a big test with them turned on and off to see if they are worth the effort in Mediocre. I am not expecting a big improvement from them in any case.
Follow the process of creating a chess engine
3 comments:
Hi Jonatan
I have not met a more productive programmer than yourself! I am so far behind :) I was wondering whether you haven't considered putting together a multi-threaded version, which will take advantage of the dual-core monsters that we currently have?
Also, your blog about says your goal is create a program that will beat yourself. Have you reached that goal yet? If not, that will be pretty scary!! :)
Multi-thread will probably take some time, mainly because I don't have such a computer myself to test it on.
I have been planning a 10 one hour game match Jonatan vs. Mediocre sometime in the near future. It would take some time though so we'll see when it is going to happen.
In blitz I get crushed every single game already, but then again I've never been a good blitz player.
I have finished the LCTII test. As I told you the previous version got an incredible amount of points (2415 for v0.3 versus 1950 for v.0.12b).
Congratulations, you manage to increase this level. The new version (v3.1) got 2455 and it found 2 additional correct answers (24 correct answers for 35 positions).
So according to this test, the new version is even stronger than the previous one. Of course the increase of points is not big, but this is normal. Now you reach a level where it will be difficult, I suppose, to improve your program significantly again, but who knows.
Post a Comment