Looks like the Deepmind Starcraft AI relied on superhuman speed after all

Originally published at: https://boingboing.net/2019/01/27/looks-like-the-deepmind-starcr.html

3 Likes

That’s not really the worst corruption of this matchup. The terrible thing is that AlphaStar used (except in the final game that it lost) a separate interface from the human which didn’t require operating camera focus.

I just wish deepmind could fund their very valuable research into reinforcement learning without over-hyping every little achievement, or allowing media to overblow the results. The whole alphastar project was worthwhile science, but the fact that they also put their fingers on the scale to ensure a media-friendly dominating result is unnecessary and disappointing.

16 Likes

Hmm. You’d have to be pretty twitchy to be able to micro your units at the top player level. The spam clicking is probably their idle-loop behavior until they need to control units with quick fire-step-fire-step movements.

6 Likes

The way deep learning works, it will always pick the least complex solution it can find. It’s always the Russian tank problem. Ignoring the obvious solution is just not what the technique is built for.

6 Likes

The Onion headline joke would surely be “Starcraft AI Hacks Human Adversary’s Twitter And Posts Nazi Flag There Hours Before Match” or somesuch

16 Likes

I think that things like that make good jokes and stories, but since some people seem to take statements like that seriously: That’s an unbelievably more complicated model.

Like the difference between:

  1. the optimal path finding algorithm in your self driving car subtly manipulating local politicians to get a by-pass built, shortening your commute, and,
  2. the optimal path finding algorithm in your self driving car deciding to drive through your neighbor’s back yard to avoid a left turn.

Deep learning is literally looking for the dumbest answer it can get away with.

10 Likes

So, if I’m reading this right, playing DeepMind is like playing a botting team. Its overall strategy sucked, but its tactics at the small unit level was what you would expect from a bot.

3 Likes

“The only winning move is…”

3 Likes

11 Likes

Once people start using deep learning for political campaigns we are going to see stuff like that, since the algorithms are likely to use some learning data from dirty campaigns and smearing your opponent by any means possible can be effective, and in the context of campaigning a simple solution.

2 Likes

AlphaGo did find a move that all established lore said was foolish in the extreme, yet it made it work and beat the best human player with it. Moments like that made me realize that these are the first steps on the road to a strong AI. Some day…

2 Likes

I love the whimsy of this article. Specially the final piece. As an old school player of the first ever Starcraft tournament on the original battle.net ladders who made it to the finals, I can attest to the click spasms the players started developing. It was akin to the moans and groans that are now common in pro tennis. It started with one woman at the Lipton in Key Biscayne and since she did well the rest of the posers followed suit even if they were mimiking behavioral byproduct instead of actual skilled moves. Now it has become the norm. In motorcycle racing we have another simmilar example with Valentino Rossi’s “Doctor’s dangle” (google it). The SC clicking phenomenon in humans is very simmilar.

4 Likes

Spam-clicking isn’t pointless. SC doesn’t use any kind of swarming AI to ensure groups of units stay together after they’ve been given a destination to move to, instead plotting an individual way-pointed path per unit based on the state of the map and position of every other unit in the game at the moment the player clicks. Spam-clicking destinations in the manner you’ll often see in pro matches helps eliminate opportunities for groups not directly observed to become splintered when something unexpected happens, like a single unit get blocked and the routing decides to send them all the way around a huge obstacle instead of waiting a tenth of a second for the blockage to clear.

4 Likes

Together with deep fake videos, we’re in for some interesting times.

4 Likes

In terms of Skynet-style robot domination, it doesn’t seem particularly relevant whether the AI simulated human limitations or not. When we’re fleeing tentacled killbots in our artfully torn earth-tone knitwear and jewellery made from old circuit boards, no one’s going to be saying “yeah, but we could easily beat them if they refrained from accurately hitting up to seventeen moving targets per second”.

Luckily, that sort of scenario is mostly silly and says more about the pundit’s psychology than about “AI”, but still.

4 Likes

Starcraft always seemed like a game that was just a contest of who could click faster, so I avoided it.

Now that an AI is trouncing people with its superhuman clickity-click I am sitting back smugly vindicated. You see, people? You see? It’s all about clicking!

Now let me get back to a Real Game: Minesweeper.

4 Likes

I’m not falling for that.

2 Likes

Making out with face masks, that’s new. And kind of pointless.

1 Like

The use of the APM average windows, rather than simple hard limits, seems like a particularly questionable choice if you want ‘human like’.

In particular, we set a maximum of 600 APMs over 5 second periods, 400 over 15 second periods, 320 over 30 second periods, and 300 over 60 second period.

Yes, it’s certainly closer than just having the bot tick along at 1000 APM for the entire duration of the match; and is probably more ‘human like’ in the weak sense that humans can likely hit a higher APM for short periods of particular concentration or familiarity than they can achieve across an entire session; but it doesn’t preclude brief bursts of absolutely ludicrious APM so long as they are accompanied by lulls to keep the window average down.

Conveniently, something like ‘preposterous micro on contact with the enemy, then either run away or advance’ is both something you can do with a second or two of burst surrounded by lulls and something that will serve you quite well in Starcraft fights.

(edit: the whole arrangement also seems like it might be the visible bit of a really annoying downside with doing black-box machine learning arrangements: those clearly work too well to be ignored, often markedly better than the ‘let’s build a nice axiomatic expert system!’ genre that has suffered so much loss of faith over time; but make it harder to provide straightforward instructions like “don’t spam-click: ‘spam click’ is any input that has no effect, either because it doesn’t target something valid(eg. ability on cooldown or is a repetition of a prior input without enough time passed for it to trigger re-pathfinding”.

Meaning is clearly overrated as a way of getting things done; but it’s really handy if you want your system to understand specific instructions to don’t-do-that-stupid-thing; which it won’t if you share none of the relevant concepts with it.)

3 Likes

I am certainly shocked that a computer is better at a computer game than a person is.

1 Like