Hacker News Reader

Comment by thefalcon | Hacker News Reader

To be fair, those time limited online games are acknowledged to be biased in favor of AI over human players, so an actual tournament style match like Lee Sedol played would be needed for a more fair comparison. That said, I expect AlphaGo to continue* undefeated (and hope to be proven wrong).

*From this point forward.

View on HN

At this point there's a basically unanimous opinion among Go professionals that humans don't stand a chance unless the time controls were skewed in a very unfair way and/or humans cooperated against the computer. Master's wins in the 60-0 games were extremely convincing (and that was using only 1 GPU).

Lee Sedol said something like a human player would need three days of thinking time vs the CPU. However it's no different than human vs car. Any car can do 300 km in 3 hours. A pro probably can do 300 km in 3 days. The different time allowance makes up for quantitative differences in hardware.

Sometimes there are also qualitative differences, no human can fly or dive very deep or go to the moon without the assistance of a machine. I don't know if there are qualitative differences in go and chess, like no human can win 60-0 against fellow pros or get to 3400 ELO despite a very large time allowance.

Total noob here, how are they biased in favor of AI?

The online matches only allowed 30 seconds per move.

I don't know if this is exactly the case for Go, but in chess, short/fast games are heavily biased towards the player who doesn't make short term tactical blunders.

In a fast game, a human doesn't have time to figure out an extremely complicated sequence of sacrifices and combinations, but a computer can look at every possible continuation of 10+ moves in the future in under a second. So not only does it not make stupid short-term blunders, but it will immediately spot any mistake the human made that is exploitable in the short term.

Up until the late 90s, and to a certain extent the early 00s, humans could use "anti-computer" strategies to win in long/slow games. A typical anti-computer strategy would be to play very conservatively and set up the board in a position that an experienced player knows has a favorable endgame, but that endgame is too deep for the computer to see, so the computer doesn't know it's being set up.

These days computers can just look 20+ moves deep every turn and have better heuristics to mostly prevent this from happening.

Go player here. Even for human-vs-human Go, fast games also widen the gap in favor of the stronger player. Although in this case the mechanism might be somewhat different - stronger players are more experienced, and their "intuitive" skills (the subconscious neural networks that tell you in a second which parts of the board look interesting) benefit from more training.

But perhaps the same reason applies to human-vs-AI Go. AlphaGo's architecture bears a striking resemblance to how the human mind operates when playing Go.

I think there might be a substantial difference between chess and go here: if you watch, e.g., AGA's reviews of the games on their youtube channel (by Michael Redmond or Jennie Shen), they make it quite clear that in the short-time-control game AlphaGo clearly outplays the human in a global, strategic sense. In fact, Redmond says the games get quite uninteresting relatively soon after AlphaGo takes the global lead, as it just tries to "wrap up". This is not about humans making blunders, like in chess, so it's seriously different from how chess engines play.

I think the main reason analogies with chess engines don't work very well is that in chess any piece can attack/capture any other piece, leading to some very intense tactics. In go, a weak group can't really attack a strong group at all.

Reply on news.ycombinator.com