Alessandro Scotti wrote:Hi François,
next thing I want to try is to take Glaurung piece/square tables and material values and put them into Hamsters, then limit evaluation just to material + PST. This should make the search very comparable.
Yes, that's an interesting experiment. You have probably figured it out already, but just in case, I should point out that the material values are included in the piece square tables for Glaurung, and that I use separate piece square tables for the middle game and the endgame, and interpolate between the two.
By the way, Glaurung and some Polish programs played a "uniform evaluation tournament" early this year. All programs used exactly the same evaluation function, a simple piece square table eval designed by Tomasz Michniewski (author of Tytan). It was quite an interesting tournament, and I would like to see more such events.
brianr wrote:Also note that Glaurung does some potentially significant evaluation scaling right at the end. In particular, beyond the midgame and endgame phase scaling it finally does this:
This may tend to minimize the "noise" of minor evaluation differences and score more positions equally, which should increase null-window cut-offs?
Yes, that's the idea. Internally in the evaluation function, Glaurung uses a value of 256 for a pawn, but the two least significant bits are shaved off before returning from the eval.
brianr wrote:Also note that Glaurung does some potentially significant evaluation scaling right at the end.
...
This may tend to minimize the "noise" of minor evaluation differences and score more positions equally, which should increase null-window cut-offs?
Indeed, I uses this technique in Kiwi and it helps there. When I ran the experiment I had modified Glaurung to test at grain size 1 too, and it did not make much difference for that position.
Tord Romstad wrote:By the way, Glaurung and some Polish programs played a "uniform evaluation tournament" early this year. All programs used exactly the same evaluation function, a simple piece square table eval designed by Tomasz Michniewski (author of Tytan). It was quite an interesting tournament, and I would like to see more such events.
Now that's a nice idea! Just out of curiosity, who was the winner?
Tord Romstad wrote:By the way, Glaurung and some Polish programs played a "uniform evaluation tournament" early this year. All programs used exactly the same evaluation function, a simple piece square table eval designed by Tomasz Michniewski (author of Tytan). It was quite an interesting tournament, and I would like to see more such events.
Now that's a nice idea! Just out of curiosity, who was the winner?
You'll find the cross table and all games on this page. The Glaurung version that played was a very early prototype of Glaurung 2, which I had started writing about one month earlier. My engine was somewhat handicapped by being the only program to play without an opening book.