|time||235 minutes plus 10/60s|
The first round started at 22:00 UTC on September 4th.
|1||AyaMC||W11R B15R W17R||B12R W1412½ B183½||B1347½ W166½||8||24||24||Winner|
|2||LeelaBot||B01R W05R B07R||W03R B16R||B12129½ W14148½ B18103½||4||32||4|
|ManyFaces1||W02R B0412½ W083½||B13R W06R||B11113½ W15122½ B1795½||4||32||4|
|4||NeuralZ05||W0347½ B066½||W02129½ B04148½ W08103½||W01113½ B05122½ W0795½||0||40||0|
Black won 10 games, White 6.
You can view any game record by clicking on its blue number in the cross-table above.
Four players registered. We welcomed a new player, NeuralZ, a DCNN program by Robert Waite. NeuralZ uses only a neural net, and spends no time on move evaluation; it still appears to be about 4-kyu. Robert has explained:
NeuralZ is a variation of the supervised-learning network described in the AlphaGo paper. It uses a convolutional neural network that is fed a view of the board and outputs a list of probabilities that a human would move at a particular position. Humans see a 2d 19x19 board and think about liberties and such. The network receives a 3d 19x19x46 board where each layer has details of the position (such as the last 8 moves, the number of liberties that each stone has, etc).
The network contains 1,771,242 32-bit floats… which is about 7MB of space. It was trained on a single 660gtx for about 4 days on GoGoD Winter (1660-present).. It reached close to 47% accuracy on guessing the one move that a professional human would make on a random board.
It makes terrible tactical mistakes because it does not use any search whatsoever. It simply is fed a board position, makes an evaluation and then takes the most confident move in the ‘greedy’ version which takes about 5ms on the host machine.
Here is a set of heatmaps displaying the network’s view of positions in a game vs. LeelaBot, similar to some of the diagrams in the AlphaGo paper. The heatmaps only display probabilities larger or equal to half a percent. Black-to-play images are what the network would do if it was the opponent, white-to-play will always move at the highest percent spot.
|ManyFaces1 vs AyaMC|
In round 2, ManyFaces1 played the move shown to the right in its game with AyaMC. This makes a second eye for the group, but in ko. Instead a move at t10 or t11 would have made an unconditional second eye.
|AyaMC vs NeuralZ|
Some Japanese players consider it impolite to win a game by too large a margin, they deliberately let their opponent catch up in the yose. And sometimes, if the opponent isn't cooperating, he has to be forced to make a good move. In its round 6 game with NeuralZ05, AyaMC did this as shown to the left, eventually winning the game by less than the komi.
Players receive points for the 2016 Annual KGS Bot Championship as follows:
|Many Faces of Go||4|