Life In 19x19

Zen (a Japanese monte carlo bot that's long been one of the top bots which recently added neural nets) beat Darkforest (Facebook's new neural net bot) to win the 9th UEC Cup computer go tournament in Japan. They both play Kobayashi Koichi 9p on 3 stones as a prize. darkforest lost, showing it still has a long way to catch up with AlphaGo. Zen is currently playing, broadcast on wbaduk. It did that typical thing of making a local corner mistake to get centre influence, will be interesting to see how Kobayashi deals with the moyo.

Edit: Kobayashi took another chunky corner in sente then easily reduced the centre so there won't be any big territories. Looks like a close-ish game. Zen leads by about 10.

In the end Zen won by 4.5 points, playing black with 3 stones. This roughly continues the pace of improvement of the top monte-carlo programs from the past few years (meanwhile AlphaGo catapults to the stratosphere):
- 1st Densei-sen 2013, CrazyStone beat Ishida Yoshio on 4h, Zen lost on 4h
- 2nd Densei-sen 2014, CrazyStone beat Yoda Norimoto on 4h, Zen lost on 4h
- 3rd Densei-sen 2015, Crazystone lost to Cho Chikiun on 3h, DolBaram won on 4h
[- Google Deepmind challenge, AlphaGo beats Lee Sedol 4-1 on even]
- 4th Densei-sen 2016, Zen beat Kobayashi Koichi on 3h, Darkforest lost on 3h.

The top 3 at this years tournament (Zen, Darkforest, CrazyStone) have all incorporated neural nets for the move selection into their monte carlo design, though I believe none have yet reproduced AlphaGo's "value network" evaluation function.

Here are the games, courtesy of computer-go.info.
First Zen:


Edit: hidden as link broken