Your latest compile for SF MCTS seems to have a problem. Under arena gui, windows 10, setting 3 min per move, I setup a position and it did not make a move even after 4 mins.
Kohflote wrote: ↑Sat Jun 01, 2024 5:02 am
Your latest compile for SF MCTS seems to have a problem. Under arena gui, windows 10, setting 3 min per move, I setup a position and it did not make a move even after 4 mins.
Guess, it's not the compile but the code, that was already known to be buggy. I just wanted to see how it would work at all at least in console, but even there it crashes now and then.
Didn't download Jim's latest version, but the one before worked best (with least crashes) in console, in Arena and in Shredder 13 single threaded (be aware, 2048Mb hash is max anyhow too), this way it ran in CuteChess too (also not at all with more than 1 thread), regards
Thanks for that too. Firefather's one yet is already of the kind of ShashChess and Brainlearn, that uses MCTS always in combination with A-B- search only, one thread has to be free for this one main line, so no "pure" MCTS- search is to be got into output except in debug- mode. The latest one with just MCTS (without A-B) I've read about so far, is Kelly Kinyama's:
BTW it's given at FireFather's github- site as example for MCTS- SF too.
I fear it's not fully bug- free neither, but already the compile from Nicolet's version you made, works well enough for me to get much information out of it, so a direct comparison to Kinyama's one would be very interesting to me too. Again bmi2 was the only compile, I'd need for my own use
