Proposal:
Entrants submit a storage device containing general purpose AI code. At the day of the challenge, each entrant's code is dumped into a computer aboard a device that is not announced until after each entrant has submitted their code. The device is situated in an environment with an obvious goal. The goal is implied by the capabilities of the device in combination with the attributes of the environment and situation. The entrant's code is expected to explore the capabilities of the device into which it is implanted by reading data streams from embedded sensors and sending control signals to actuators (arms, wheels, eye focusing and pointing, etc.).
The challenge device may be a submarine, a car, a microscope, a motorcycle, a missile, a fish, an airplane, dirigible, or helicopter, a walking or crawling robot, a kitchen robot, a bank ATM, a database, etc... there may actually be two or three devices that each entrant's code must inhabit and learn to control in a situation with an implied goal and specific dangers.
The idea is that AI challenges have, to date, been too specific and narrow leading to brittle customized "AI" that is in point of fact only tricky logic and programming that is not intelligent at all (can't learn, adapt, or improvise).
The entrants can assume that the challenge device(s) will deliver several sensor data streams and output ports for the control of actuators. More detailed input/output protocols and conventions will be published well before the challenge date.
The winner of each challenge will be selected through a combination of subjective judgement by a panel of experts and objective metrics (exp. time-to-completion vs. mips of processing used).
No comments:
Post a Comment