Computer Go - Strength and Weakness

7 posts / 0 new
Last post
macelee
Computer Go - Strength and Weakness

I've seen games between a human player and Zenith Go, a strong computer Go engine. The games are quite interesting and show clearly the strength and weakness of the computer AI. The program plays reasonable opening (I suppose this is not hard to achieve because of the highly advanced pattern match algorithms and large professional game databases available). Its reading ability locally is quite good and rarely makes low-level mistakes. Surprisingly its yose appears to be quite weak - I would like to think yose is something that can be quantified more easily so computers should be good at it.

But the biggest weakness I can see is that Zenith Go rarely plan anything globally, nor can it understand its opponent's global plan. I can explain this using a real game.


How computer AI is tricked

Black (the computer AI) is having big potential in the middle of the board. Apparently there is some aji around 'a' but jumping directly into the center at White 6 doesn't work (note that 'b' is black's sente so the group is quite safe). So White square wanted to prepare for the invasion. This is a very natural thinking by human-beings but computers must have a lot of difficulties understanding this. Black 1 is a sharp move locally, but clearly the computer AI doesn't realise the danger elsewhere. White 2 continues his plan and white is tricked into playing the cut. Now White 6 can easily break into the center.

macelee
Good reading ability

Now black's centre was completely destroyed. The human player was so happy that he played too quick and made a mistake


Accurate reading

White only need to secure the weak group to win the game comfortably. But White 1 and White 3 were wrong. The computer clearly read this, giving up 3 stones in exchange to cutting off the big tail! Well done! (BTW, white can give up two stone and connect the tail, the dragon won't die)

macelee
Computer tricked again

Well into the yose stage but something interesting happend.


Tricked again

They are fighting a small ko. White make *w1*/*b2* exchange as ko threat, knowing that black cannot play *b4*. But the computer AI played Black 4 immediately. Apparently the computer did not even consider the marked stone (*bs*) are part of the calculation.

Tony Hiran
Yose

I'm very surprised that the AI got tricked in the yose. With its good reading and counting ability I would expect it to have also calculated that it would have lost more points by playing that than it would have gained...

moboy781
How strong was Zenith Go?

How strong was Zenith Go?

macelee
Re: How strong was Zenith Go?

I understand Zenith Go is the commercialised version of Zen so it is definitely weaker than the robot Zen player on KGS. I estimate it is at a low dan level.

Metallian
Metallian's picture
Software improvement

After reading this and about Ishida Yoshio's defeat by Crazy Stone, I wonder if we're getting to the stage in which a Go playing program will defeat consistently any professional.
What do you think?