somebody asked me to post my views on 9x9 go on the list based on my experience with correspondence go on OGS and little Golem.

I have been playing online correspondence tournaments since 2011 with Valkyria which is a MCTS heavy playout MCTS using AMAF heavily tuned for 9x9. Also with the kind Support of Ingo, I used to generate a lot of 9x9 data for opening book preparations many years ago.


During these years I collected a opening book based on running Valkyria with 2 thread something like 2 to 24 hours. I can do this because of a hash table that works well. Valkyria is tuned to be very selective, so it follows an iterative deepening algorithm where it searches for some time and the discard the tree, storing the best move. In each iteration it will start a new search with an empty tree, but will use the hash table to research known position more efficiently. This way it can overcome the problem that a MCTS fills memory very quickly.

Valkyria has no stopping rule, so in the end it is mostly a hybrid human/computer decision when to stop search. But most of the time I just wait until it seems to converge on a single move with a clear winrate advantage. For the openings I tend to do choose moves many times, mostly to avoid lines where it has lost in the past, but I only choose moves it has been investigated during iterative search.

If I run Valkyria on 9x9 CGOS (2295 Bayes Elo) it is not very strong, but against amateur humans on OGS and LG it has been very successful but not unbeatable.

On LG there is a player Gerhard Knop (who I think uses one or more programs as support (or at least used to do I just read this indirectly somewhere) which seems to be clearly strong right know. At least recently he seems to be very good with white against Valkyria.

So what have I found out about 9x9?

I used to think that with the Opening book of Valkyria black is an easy win with a komi less than 7.0. Since Gerhard Knop has been beating Valkyria with white I changed into thinking black should be an easy win... but my opening book is not very close to optimal play, it is just the playing style of Valkyria.

Other human players are playing very well too and it often happens that Valkyria wins games which was evaluated as a loss despite the enormous (well for a single PC guy, I am not Deepmind) computations behind all moves. The strongest humans repeatedly play moves that Valkyria never read deeply, even after 12 hours of computations, which turn out to be as strong or better than the expected best move.

So is 9x9 easier than 19x19? Yes of course... but it is not that easy. In go there is the complexity of the number of legal moves but this is no longer the big problem. Most moves can be searched safely very shallow or not at all. In a well played 9x9 games it is a simultaneous problem of: endgame, life and death, semeai with ko fights, subtle differences in move ordering for forced moves etc. This give fighting lines that cannot be reliably evaluated by MCTS until read 40-70 ply deep because all stones are unstable.

I have not yet trained a value network for 9x9 but I can imagine that it might still be very hard to get close to perfect evaluation, so any engine would still need to search very deep to play close to perfection.

From the current surge of strong engines on CGOS 9x9 I just learned that my engines is even further from perfect play that I previously thought since there are no sign of these engines being near perfect play given win/loss/jigo statistics.

I did 9x9 computer go for many years and I think 9x9 go is much harder than I originally thought. I am not ruling out a super strong 9x9 go program appearing next weak. I am just saying that close to perfect is much stronger than that I have seen so far.

Magnus Persson
Computer-go mailing list

Reply via email to