What Else Can AlphaStar Do? Or, What Can You Do With Transformer Neural Networks?
tl;dr—You could call it “practical intelligence” in terms of the kinds of high level tasks it can learn. Practical is often good enough, for drones and robots.
AlphaStar is the A.I. that plays a complex, real-time strategy game called StarCraft 2, winning against the best in the world. It is developed by Google’s Deep Mind. That’s neat, but how can this technology be used? This is your cybernetic report.
AlphaStar is designed to emulate a human playing on the computer, and as such the AI’s input is pixels. That is, AlphaStar sees roughly the same screen a human player sees, as input.
So right away we have on some real level an AI good enough to use modern computer technology. Any job that is a closed complex of purely point and click activity can be performed by this A.I., done deal.
The typical “value-add” to that kind of work, the human part, is information gathering and synthesis. Incidentally, the underlying neural network AlphaStar uses, called a Transformer, is also very good at translating language. This is a strong basis for a understanding information on a useful level. Add some other functions to the mix, and watch a lot of the information economy become fully automated.
AlphaStar isn’t going to use spreadsheets or graphics editors, yet. However, there are other neural networks have been developed that can take as input descriptions — “dog wearing glasses” — and produce that image. The same networks can be used to transfer style, so now “dog wearing glasses” was painted by your favorite painter or pop artist. Build a graphic editor out of these generative neural networks — infinitely modulateable convolutions — and let AlphaStar play with it. Now you have at the very least a very deft creative assistant.
Yet, that is but the superficial analysis of AlphaStar. Just look at what it’s really doing, when it plays this game. AlphaStar is managing resource gathering, serializing tasks for agents, and contriving battle tactics, in a real-time, totally dynamic environment (against a live human opponent). AlphaStar is bossing around bots, in sometimes clever and crafty ways. If it can beat a pro gamer, it can beat the pants of most any management.
If you have robots can move around, lift and lower, grab and release, or whatever they can do, they have a managerial oversight brain-center in AlphaStar.
Or, even if your don’t have bots, maybe you still have human factory workers, AlphaStar can tell them what to do, too, probably better than any alternative. If the system is not life or death, as it is in space battle, perhaps AlphaStar is good enough already for the job.
AlphaStar can control drones in a complex and dynamic environment. That environment could be farms, factories, oceans, or asteroids. It could be cities, delivering mail, meals, packages, and supplies. AlphaStar doesn’t train robots to perform their duties, it doesn’t actuate it’s movements. AlphaStar sets the vectors.
A.I. is also being used to teach drones how to move, and do fine-tuned things like pick up fruits, and actuate hand-like machines. It would be only a mild surprise to find out Transformer neural networks also performed well at that. If you can model the environment as a transformation, who knows what you can do.
Here’s a quick human-helper scenario. You’re at the Louvre, a gigantic museum impossible to view in one visit, and full of dusty stuff you don’t want to see anyhow. You only want to see you favorite styles, artists, and epochs, and miss the rest. So you give AlphaStar’s cousin this directive, and it tells you where to go from, one place to another, all the while figuring out which is the best way for you to “beat” traffic around you even as it evolves.
Wait, you say, this is maps and directions. That is correct. AlphaStar shows that transformer networks are versatile. It is likely that many things can be done with them in properly modeled reinforcement learning environments.