Cyber police have collected a set of pictures of “friendly” and “unfriendly” robots, and asked their analysts to write rules that tell the two apart based on visible features — the shape of the head, jacket color, height, antenna color, what’s held in each hand, whether the robot is smiling, and so on. Simple rules that worked for the first batch of robots failed badly once a second, larger batch of robot images came in.
The challenge: given the two sets of labeled example robots, come up with a rule set (or a learned model, or some combination of the two) that reliably classifies both the original robots and new ones as friendly or unfriendly — and ideally can keep adjusting as further examples arrive. (The scenario and images were adapted from a 2008 machine-learning research paper on rule extraction.)
Send your solutions to DecisionManagementCommunity@gmail.com, or open a pull request to add yours here.