That's pretty interesting.
The part I find particularly interesting is this bit:
the AIs largely worked out how to tackle the problem themselves. “They’ve come up with protocols that are different from how humans solve these problems,” says Foerster. “We don’t yet fully understand what the solutions are, but we know that they work.”
This is something I see pretty often in AI projects. Â The solutions that AI's come up with are novel and often non-intuitive…if we can even understand the solution.
It reminds me of this thing I've shared before: Â http://goo.gl/Dkl5WQ
Basically, starting from a random design an evolutionary algorithm designed a computer chip for distinguishing tones. Â After 4000 generations, the chip could distinguish between a 1khz and a 10khz tone.
…no one had the foggiest notion how it worked.
Dr. Thompson peered inside his perfect offspring to gain insight into its methods, but what he found inside was baffling. The plucky chip was utilizing only thirty-seven of its one hundred logic gates, and most of them were arranged in a curious collection of feedback loops. Five individual logic cells were functionally disconnected from the rest— with no pathways that would allow them to influence the output— yet when the researcher disabled any one of them the chip lost its ability to discriminate the tones. Furthermore, the final program did not work reliably when it was loaded onto other FPGAs of the same type.
The relevant part of this story is that there were parts of the chip that served no apparent purpose, and in fact, were totally disconnected from the rest of the chip. Â However, if the research disabled these parts, the chip stopped working.
It seems as if the operation of the chip became dependent upon these disconnected parts of the chip not through an electrical connection but (possibly) through the magnetic fields caused by electrons traveling in them.
