Saturday, February 9, 2019

Star of Alphas

AlphaStar is the newest and best Starcraft 2 AI that is attempting to end human dominance in the game forever.  It is built by the same people who made AlphaGo, that finally ended humans as the dominant force in Go, so expectations were high.


AlphaStar isn't quite there yet.  It only knows how to play Protoss vs. Protoss, and only on one map.  In a real tournament setting it would get massacred - basically it wouldn't know how to play.  But in that limited setting it beat LiquidTLO 5-0, which initially people thought wasn't such a huge deal - after all, TLO is a Zerg specialist, not Protoss.  But then it beat LiquidMana 5-0, and Mana is one of the best Protoss in the world.  Finally it lost a single game against Mana later on, once the team put some extra restrictions in place on AlphaStar.

The interesting thing is not the stuff that AlphaStar could do that no human could do.  It could micro in three places at once, and used that superior control and reaction time to smash the experts, no doubt, but we all knew that would happen.  Even though they limited its actions per second to mimic the maximum a pro can deliver, its efficacy with those actions was far superior.  The cool thing was what it did that humans *could* do, but don't.  It built a ton of extra probes in a way that pros simply do not do.  It also was cavalier about harassment early on, and seemed to not care about losing early economy so long as it got to build a perfect push and smash in the opponent's front gate.

It is hard to say if building those extra seemingly pointless probes was tech that AlphaStar has figured out that people haven't, or if it is a thing that only AlphaStar can capitalize on.  Perhaps it is just a way to compensate for it's weakness against harassment.  Hard to say, but when the commentators lose their minds over a play and make it clear that a human doing this would be laughed at, but then that player wins a crushing victory, you gotta pay attention.

This is much like it was around 2000 when Deep Blue was finally beating Kasparov at chess.  The machines haven't quite won yet and they need optimum conditions to take the best humans down.  However, it is totally crystal clear that in the next couple of years algorithms will improve, computers will quadruple in power, and humans will forever be second best.  Someone suggested that this will change how we think about Starcraft 2 competitions, but I think that is silly.  We didn't stop racing the 100 meter dash when we invented motorcycles, or substitute pitching machines for pitchers in baseball.  This won't change human competition at all.

I suspect that just like in chess people will train themselves against AIs to test out builds and theories. Everyone will accept that computers can't be beaten, but we will still be interested in watching the human personalities battle each other, just as we are in chess championships.

I love these moments where the tipping point is reached.  For decades AIs have been pathetic in Starcraft, obviously inferior, never able to compete.  For the rest of civilization AIs will be assumed to be unbeatable.  But today, right now, we are at a point where the AI is just on the edge, able to crush humans on certain days, in certain ways.  These moments are wonderful and fleeting and often teach us important things about what we have always taken for granted.

No comments:

Post a Comment