Phase-Functioned Neural Networks for Character Control [video]

Did anyone else play with the BEAM robotics circuits back in the early 2000s? They were an excellent introduction to both the basics of artificial neurons, and to the analog side of digital circuits.

Neurons were constructed by passing an inverter's input through an RC high-pass filter. A ring of these would slowly feed a single pulse, or a pattern of pulses, around the network by something equivalent to phase-shifting an incoming pulse. Sensors, and current or voltage feedback from a straining motor, could adjust the timing or even alter the pattern of the pulses. The pulse train would directly power the movements of the robot.

http://solarbotics.net/bftgu/starting_nvnet.html

Absolutely jaw-dropping. I can't wait to see this implemented in any of the upcoming games.

The paper: http://theorangeduck.com/media/uploads/other_stuff/phasefunc...

Would this add much overhead to a game that may already be taxing the system?

From what they claim, no.

"Once trained, our system is also extremely fast and compact, requiring only milliseconds of execution time and a few megabytes of memory, even when trained on gigabytes of motion data."