3 ms·
I'm always partial to a good argument for being sceptical of such announcements (I'm a professional skeptic; "researchers", some people call us), but in this ca
by YeGoblynQueenne 8y ago
I'm always partial to a good argument for being sceptical of such announcements (I'm a professional skeptic; "researchers", some people call us), but in this case the question is whether it is ever possible to make the comparison fair at all. For example, we can restrict the AI player to perceive the game world in the same way (iiish) as the human player, but we can't really make the human player see the game world in the same way as the AI player. So if the AI player has an inherent advantage because of its superior perception, there's just no way to level the field and compare the human to the AI in a fair manner. Perhaps we just have to accept that a human will never have an advantage against a computer player in a computer game, or that computer games are just not very interesting testbeds for AI, after all (that would be very controversial).
On the other hand, it's also useful to remember that previous attempts have been made to develop strong Starcraft players, and they all had access to the same API as the DeepMind player, yet they didn't perform as well. Of course it makes a difference that it's DeepMind who achieved this and they seem to have spared no costs in hardware and training time, unavailable to a smaller team. Perhaps we shouldn't be that surprised to learn that a poweful computer can do better than a human in some cases. But, at the end of the day, the result they got against MaNa shows that it is possible for an AI player to beat a strong human player -fairly or unfairly- in a game that was rightly considered very hard for AI players, under any assumption.
- kilburn 8y agoA fair comparison is simple: make the AI play the game using a standard screen, mouse and keyboard (the game was designed to be played this way after all, and non-standard peripherals are not allowed in tournaments by the way). It suddenly seems WAY more challenging for the machine, doesn't it? You could also invent some contraption that interfaces directly with the human's brain to even out on the other front, but this sounds even more futuristic. The average APM limitation is just a publicity stunt. There are similar caps one could apply that would make the challenge much more "fairer". For instance: Max (instead of average) APM or limiting mouse travel distances (no one can get even remotely close to clicking 300 times/minute on alternate sides of the screen with pixel-perfect accuracy AND timing). Finally, the AI only plays 1 of 6 possible match-ups in 1 of infinity possible maps, and the live game has shown severe strategic mistakes (mainly the indecision when getting harassed and the inability to build a phoenix to counter the prism / keep harassing itself with the oracles instead of stupidly watching over the prism with them). It still has a long way to go for me to consider it "superior" to humans regarding intelligence (not just superhuman control). All in all, this is still an impressive achievement though!