Time Control in AlphaGo's Recent Games

General Go Discussion
1

I’ve read that the time control in AlphaGo’s (AG’s) recent games was usually 3 x 30s byo-yomi (or was it 5x, I’ve forgotten) with no maintime, and a few games were 1m intervals. This is very fast play, and computers today have a distinct advantage at speedy play. Even so, some of the games seemed pretty close. So the question arises, would AG have done as well if the control had followed normal tournament or match standards? And if not, how many games might it have lost? Or am I grasping at straws?

2

The problem is the games that seem close might not have been. As AlphaGo’s estimated win % grows, it plays more and more conservative, even to the point of playing “dumb” moves no human would play. But if that dumb move better guarantees a victory by .5 vs a normal move being a bit riskier but a possible win with 5 points… it’ll take the .5 win. Since AlphaGo only cares about W vs Loss, not score.

So I’m hesitant to ever put stock in a “close” game vs AlphaGo. :slight_smile:

1 Like
3

But what about the time control?

4

Honestly I just don’t know. I’m too much of an amateur myself. :slight_smile: I’m just cautious about thinking any AlphaGo win was close…

5

Not sure I fully understand what you mean, but if I do, I would say that since the previous version that played against Lee Sedol (which were 2 hours for each + byo-yomi) won quite convincingly I don’t think it would have mattered (especially if this versionis is superior). If anything (and this is just my oponion not based on any deep undrstanding) I would say quick play would favour the human. All in all we can play on quite a good instinct, while computer has to calculate the variations and that takes time even in this century. And can easily just not get to that one variation that is important.

I would be very interested if they improved Alphago’s time management though. As I understand it it was one of the losing factors in that only game lost by alphago against Lee Sedol. Where he played an unexpected and complicated move and alphago responded after a couple of minutes as always, but with an inferior response, while the commentator (Michael Redmond 9p) explained that if this were his game he would gladly sacrifice the rest of the clock (alphago had plenty of time left) to thoroughly explore all the variations.

1 Like