TalkChess.com

[Vincent Diepeveen]

"I don't expect it would work that well for chess."

I agree with the above.

Over my dead body it can beat todays selective searchers.

"But if I were to go back to programming chess, I would try it."

This is total bullshit of course.
You are someone who likes to win bigtime.
I saw it in your eyes several times at world championships.
No freaking way to deny it.

You are NOT going to implement in Le Fous Fous (the crazy bishop) something that is in advance going to lose for you.

You hate losing more than any other game tree addicted nerd in France.
Don't deny the obvious please.

It just works for your go program, because your go program is THAT bad compared to objective human standards, and you didn't figure out a normal way to search in go yet.

As soon as go programs can play tactical without mistakes, then all this monte carlo nonsense (nonsense from game tree playing programs viewpoint seen) dies slowly as it keeps losing more and more against tactical more bugfree playing programs.

Currently search in go gets dominated by hard forward pruning in root, and using internal zobrist keys for repetition of 32 bits integers.

What do i need to say more to prove how infantile search is currently in go programs with respect to "not having a worst case by means of superdubious pruning".

In the land of the blind, one-eye is king (dutch saying).

For go it is possible to prune real accurately. Just use this search form:
in each given position divide all moves in 7 groups.

Group 1 is the seemingly best moves you want to try.
Group 2 is a bit less great.

and so on.

Now instead of the normal reduction of 1 on each position after making a move as we have in brute force search, you reduce simply the group number where the move belongs to from your search depth left.

In Qsearch you look first 4 ply to all moves that shut in groups that have 2 liberties or less. You extend infinitely all lines that close groups which have 1 liberty.

Now just make some go knowledge classifying those 7 groups and kick the hell out of your montecarlo method with it.

Best regards,
Vincent