Best engine without monte carlo playouts or AlphaGo methods?

[Slarty]

What's the "state of the art for traditional methods" in computer Go? Have there been any advancements since the application of monte carlo playouts, or would it still be a big deal to come up with even a shodan bot (or efficient life and death solver) that doesn't use them or networks trained from human games or self-play? This has been my dream. Hi, and thank you.

[Mike Novack]

Prior to introduction of Monte Carlo methods, the best of the rule based programs were at about 6 kyu.

a) And THAT level of play including :randomize" so not always responding in exactly the same move.
b) And at least one of these could be asked "why" (what were the go principles used to choose that move)
c) Probably playing too passively.

That said, for a person learning go, up to say maybe 8 kyu, such programs might still be useful learning tools, especially if recording game and then replaying asking "why"

Could they be made stronger? Possibly given what has been learned from the AI programs in terms of things we thought we knew but were wrong, the rule based programs could be made a few stones stronger <<nothing is learned from Monte Carlo >> In other words, stronger than what they are now but unlikely up to 1 dan amateur.

[Slarty]

Rules-based approaches are one thing. I maintain we're missing an algorithm and a couple heuristics, because humans don't play go with complete playouts or learn it from (millions of) silly plays.

I have a lot of grudging respect for MCTS despite taking opportunities to call it throwing spaghetti at a wall. It teaches some things, probably.

[gennan]

Perhaps the (open source) GNU Go engine is the strongest engine before MCTS. The latest version is from 2009 and it may be about 6k KGS.

Some more info about GNU go: https://senseis.xmp.net/?GNUGo

[Mike Novack]

Perhaps the (open source) GNU Go engine is the strongest engine before MCTS. The latest version is from 2009 and it may be about 6k KGS.

Do not forget the (non-open source) Many Faces of Go and Go++. They were also that strength. Look at the computer go tournaments of that period.

Note also that in the transition to Monte-Carlo programs (Many Faces succeeded with that, go++ did not*) while MFOG not the strongest of the MonteCarlo programs** still a useful teacher to around 1 dan. It could give the go principles represented by its choice -- it was using Monte-Carlo to choose between rule based candidates. In other words, could NOT say why one set of principles outweighed the others (Monte-Carlo decision) but you could see which principles involved

* I was beta testing for that attempt

** That, however made it weaker than the best pure Monte-Carlo programs. SOMETIMES the best move (discovered by Monte-Carlo) not among its candidates.

[And]

Handtalk 5 plays about 1 kyu
https://lifein19x19.com/viewtopic.php?p=275034#p275034
It's a very interesting program