New networks Leela Zero, a Sensible Artificial Intelligence

[And]

http://sai.unich.it

[Vargo]

Very nice, thanks !

sai9x9 seems very strong, about as strong as katago1.2 ?
9x9 Sabaki games between katago and sai are often very close.
(here, both programs play around 2-5 sec/move)

sai9x9 can win against gnugo at H2 and H3, and even sometimes at H4 (komi 0.5)

[And]

if I understood correctly, sai9x9 uses a network 94619dea? but the network was received on 12/21/2018? almost a year ago! but strange, I haven’t seen this network before

[And]

last network 848fe9be (10 sec/move, cpu version) http://sai.unich.it/networks/848fe9be1b ... 99fbf84.gz
against zen 7 1 dan (gtp4zen -t4 -s1800 -n1 -o1.3 -p0.35) - zen 7 surrendered on move 109

[And]

last network 43268842 (10 sec/move, cpu version) http://sai.unich.it/networks/43268842e5 ... 5d364fe.gz
against ZEN 7 2 dan (gtp4zen -t4 -s2000 -n1 -o1.0 -p0.4) - SAI won +2.5 (302 moves)
20 more games SAI & ZEN - 4:16

[And]

SAI #64 fe1cfa3c (10 sec/move, cpu version) http://sai.unich.it/networks/fe1cfa3c81 ... 95f2dfb.gz
against ZEN 7 3 dan (gtp4zen -t4 -s2200 -n2 -o2.0 -p0.45) - 2:18
zen surrendered in two games:

[Vargo]

Two more Sabaki 9x9 games Sai-Katago (both at 40000 playouts)
Maybe Katago 1.2 wins slightly more often than Sai

[Vargo]

Two Sabaki 9x9 games Sai-gtp4zen (40000 playouts for Sai, -T10 -s40000 for Zen)
Sai wins even games and can win H2 with komi 7.5 (but not always)

[frmor]

Hi, SAI main author here!

What options are you using with SAI?

Be aware at least of the following ones:

--lambda [0.5]
--mu [0]
--nrsymm

The latter is the simplest. Merge symmetrical positions in order to explore more deeply. Highly recommended when playing at fixed visits as it could give quite a lot of strength increase

lambda and mu are real parameters intended as 0<=mu<=lambda<=1. They tell SAI to consider a range of komi values when playing, instead of just the real komi of the game. If you are playing with an opponent of equal strength it is recommended to set both to 0 (notice that this is NOT the default for lambda).

If you are playing with a weaker opponent, and/or there is handicap, the default values should be ok, at least for 9x9

[And]

frmor, thank you so much for this project!
I correctly understood, these recommendations for the case if SAI are used to play against another program (human)?
which values are better if the opponent is stronger? and mu = 0.5 or 1 can be used, and what will it give?

[Vargo]

--lambda [0.5]
--mu [0]
--nrsymm

and --noponder -p 40000 --komi x.x
Thank you for your detailed explanation about lambda and mu. I like SAI very much

[frmor]

frmor, thank you so much for this project!
I correctly understood, these recommendations for the case if SAI are used to play against another program (human)?
which values are better if the opponent is stronger? and mu = 0.5 or 1 can be used, and what will it give?

It's a bit complicated, and you should read the paper to get full details. Let us say that lambda and mu control the two sides of the interval of komi values over which the winrate is averaged.

If mu=0 it means to keep the real komi as one side.

If lambda=1 it means to use, for the other side, the komi value that SAI thinks would make the current position fair (in the sense of 50% winrate).

Intermediate values are to be intended in the winrate scale, not komi scale, so that for example, if SAI thinks it is ahead, with winrate 80% and lambda=0.3, than it would compute the weighted mean 71% = 0.3 * 50% + 0.7 * 80%, look for the komi value that would give itself a 71% winrate, and then use this as the side.

So if SAI is black and real komi is 7.5, then this virtual komi value would be something more, say 9.3. Then SAI for the exploration of the current move would compute the nodes winrates as averages over the komi range [7.5 ; 9.3] and choose its move accordingly.

You can put mu=lambda>0 and in this way SAI would play as if the komi was the one computed, hence giving itself increasing score handicap when ahead. (And the opposite when behind.)



My recommendation is to use lambda=mu=0 when against a strong opponent, to try hard to win with the real komi.

You should use values larger than 0 when against a weaker opponent, both to improve the final score, when playing a fair game, and to improve the strength of play when starting with handicap.

When SAI is behind with a weaker opponent because of handicap, positive values of lambda (and maybe mu) were showed to improve its ability to recover from the initial disadvantage.

[And]

SAI #70 c74e6c7d (10 sec/move, cpu version) http://sai.unich.it/networks/c74e6c7d84 ... 2c385ee.gz
against ZEN 7 4 dan (gtp4zen -t4 -s2400 -n2 -o1.5 -p0.5) - 1:19
zen surrendered in one game:

[And]

SAI 3ca44219 (10 sec/move, cpu version) http://sai.unich.it/networks/3ca44219f0 ... cd19998.gz
against ZEN 7 5 dan (gtp4zen -t4 -s2700 -n2 -o1.0 -p0.55) - zen surrendered:

[splee99]

Zen behaved weird in the last few moves, which seem to be low level blunders.