Discussing AlphaGo’s style: Ke Jie and Fan Hui

In a three-part video series released by DeepMind, Ke Jie and DeepMind’s Go Ambassador Fan Hui review the AlphaGo vs Ke Jie games.

Here I have paraphrased statements about the AlphaGo style of play. I have combined both Ke Jie’s and Fan Hui’s statements; I have also combined related statements from different parts of the videos.

On the opening and close combat:

In the opening, many moves are possible. They often only differ by one or two percent at most. In the opening, you can really play whatever you want. Nothing is impossible. But in close fights, if it doesn’t play a certain move, something could die at once. The largest gap between me [Ke Jie] and AlphaGo lies in close combat. It always finds the crucial move. I had no idea how it would approach this [position]. But it had already evaluated everything.

On copying AlphaGo’s moves:

Its opening moves such as this attachment are easier for us to learn. But fighting is a whole other story. We can’t simply learn to read as far as it does. We can only rely on our own reading and skills. In Go, we don’t usually get an identical position again. If something is altered ever so slightly, copying AlphaGo’s moves could prove to be fatal. It wouldn’t play the same move itself. So we can only take ideas from the opening. I think what we can learn is its way of thinking. It’s useless to learn specific moves. But this attachment [to a 4-4 stone in the opening] does not take reading skills. It’s more about the direction.

On joseki:

We need to be careful mocking other people’s moves. AlphaGo might like it. I [Fan Hui] stopped teaching my students joseki. They can play whatever. But in fights I can point out problems in direction, reading or shape. There’s no such thing as “joseki”. Joseki is not important. We should learn to utilise our stones to their utmost efficiency.

On win rates:

It can say that one side has a high winning percentage, but that could mean a difference of only a few points. A small mistake can reverse that. A win rate of 60% does not decide the game. It would not slack off at just 60%. The opponent would not start playing with fire either at 40%. A win rate of 60% is not really a lead. It’s still close. The advantage is so tiny. […] Even 70% is not yet the point of no return. A win rate of 80% is untouchable.

On playing normal moves even when winning:

Even if it has an 80% win rate it does not play slack moves. It doesn’t matter whether it’s winning or not, it’s still searching for the tightest move. It doesn’t get scared into becoming defensive. It still plays what it should. That’s completely different from the previous version. Formerly, it used to retreat when it was leading. When it played Lee Sedol, I sensed that it would yield a little when leading. Perhaps its understanding of an advantage has changed. When there are still many unfixed places and there are still, say, 80-90 moves left, it does not slack off. It plays normally. It has no concept of slack or aggressive moves. It just knows that it should keep playing the tightest move to raise its win rate. That’s it. Because it if loosens up, its win rate might drop and it might lose.

On efficiency:

All its stones are lively and working together. They are well-placed and useful. It never has unnecessary stones. Go is a game between stones and maximising their efficiency. AlphaGo always knows the follow-up to its moves. Us mortals only see one step ahead. AlphaGo not only sees ahead when it plays, it also takes its previous stones into account. It’s “environmentally friendly”.

We should make our move raise the value of our other stones that we played before. Then the value of our move will also increase. Kind of like team spirit. You are the team captain who organises your team members. You do not develop one move but your entire team.

On the balance of the strength of stones:

Unlike us, it considers the balance of the strength of stones. That’s way too difficult for us humans. We don’t know what’s weak or strong. In its self-play games too, the strength of groups is constantly changing. Sometimes, a seemingly thick group may suddenly die. That’s really bizarre. You see strong stones, and after two moves they become weak. The first impression can be misleading. You think it’s strong but in reality it’s not. It’s difficult to judge objectively.

On being objective:

AlphaGo is always objective. It’s unaffected by emotions. Losing doesn’t matter. Just play again. It knows no concept of good or bad. It’s only shapes and efficiency. We can learn a lot from that. It always searches for the best moves. Unlike me, who after a mistake gets in a bad mood. I try to play excessive moves to make up the loss. AlphaGo is different. It just keeps playing the best moves it finds. AlphaGo is still far from perfect play. It does make mistakes. If you calm down, you might still have a chance. That’s on a higher level than trying excessive moves. It’s always calm. That’s a question of confidence. Once you lose confidence and you fail to calm down, you go all-in. But that could be more dangerous.

On confidence:

Facing AlphaGo, [feeling confident] is indeed difficult. You cannot sense its emotional state; it’s easy to start doubting yourself. When it played in Seoul with Lee Sedol, in the second game it played the famous fifth-line shoulder hit. Crawling on the fourth line was an easy move. But Lee Sedol decided to push up. He probably thought that if he crawls, Black will jump or something and he would lose control over the game. That’s why he pushed to try to initiate a fight. That’s an emotion thing. Facing AlphaGo, it’s easy to lose track of your emotions and start doubting yourself, believing you’ve made a mistake.

It’s hard to humans to recover from such a blow. Especially against AlphaGo. It would be different facing a human. You lose hope. You can’t catch up after a mistake. Against a human, you’d calm down and wait for a chance. Facing AlphaGo, you’d worry that there is no more chance. But that’s not a worry, that’s reality. There is really no chance.

On whole-board direction:

After local exchanges it plays somewhere else. When stones cluster together, that area has already become smaller. Directionally speaking, it makes the exchange into a sort of probing move. After finding out how the opponent responds, it switches to the bigger direction.

That’s a weakness of us humans. It looks like Black has to respond [to this invasion]. How could Black not respond? [But maybe you don’t know how to answer the invasion.] Often if you don’t know how to play, it’s better to play nothing. That’s typical for AlphaGo. It won’t scuffle over local details. It plays whatever is important.