Robotics Scientists Decide Robots Need to be Nurtured

The very interesting article For robust robots, let them be babies first reveals that learning machines are capable of developing algorithmic improvements faster than humans can program new capabilities into the algorithms. The somewhat vague description of how these specific robotic systems work leads me to wonder if swarm theory doesn’t play a greater role in robotic intelligence than we have previously believed.

It makes sense on a certain level. In the 1980s computer scientists were concerned about President Reagan’s Strategic Defense Initiative because it called for tracking multiple targets in real-time — a capability that computers lacked at the time. Although that technology exists now, it occurs in high distributed environments. Concurrent programming models in the 1980s were rather primitive and computer scientists could not prove their algorithms were correct.

The correctness of an algorithm ensures that it will behave as expected and desired under all circumstances. The last thing you want in a nuclear war is for your missiles to turn around and aim themselves at your own cities — or for your defense systems to misidentify your missiles as their targets. Object-oriented programming allowed software to spin off new copies of precoded routines (this is called instantiation — meaning a new instance of an object is spawned) but the code did not evolve.

In fact, if you’re trying to prove the correctness of an algorithm, the last thing you want is for the algorithm to change itself. Self-modifying programs have been around for decades. Learning self-modifying programs are extremely complex and unpredictable. Swarm theory might provide some resilience in a highly chaotic environment like that, or maybe we should say it might establish some boundaries that the autonomous self-modifying algorithms could respect.

By creating a lot of little algorithms that cannot do much by themselves, but which learn to work together and slightly improve their abilities to work together, we might be able to create some self-learning artificial intelligences that observe certain collective rule sets. Change and growth could follow more-or-less predictable patterns, even though we may not yet have all the math we need to assure ourselves of those predictions.

Maybe this is what the robotics team is doing. Maybe I’m overthinking the process. Still, it makes sense — when confronted by massive challenges — to break problem sets down into smaller sets that are more easily solved.

Our ultimate goal is to create a form of artificial life that does not desire to control its own destiny. This artificial life would provide us with a limitless labor force capable of building structures and societies we can only just begin to imagine today.

That is still mostly the stuff of science fiction and fantasy.