Researchers Mark Riedl and Brent Harrison from the School of Interactive Computing at the Georgia Institute of Technology have proposed a technique called Quixote that teaches "value alignment" to robots by training them to read stories and learn acceptable sequences of events. The work builds on Riedl's earlier Scheherazade system, which gathered correct action sequences by crowdsourcing story plots from the Internet.
"The collected stories of different cultures teach children how to behave in socially acceptable ways with examples of proper and improper behavior in fables, novels and other literature," Riedl said. "We believe story comprehension in robots can eliminate psychotic-appearing behavior and reinforce choices that won't harm humans and still achieve the intended purpose."
In their paper, Riedl and Harrison describe how Scheherazade first teaches what constitutes a normal or "correct" plot graph. Quixote then translates that graph into a "reward signal" that reinforces certain behaviors and punishes others during trial-and-error learning. The technique maps potential unethical actions an agent might take, allowing future researchers to adjust accordingly.
A hypothetical example illustrates the approach: a robot tasked with picking up emergency over-the-counter medication as quickly as possible could choose to take the medicine without paying, interact politely with the pharmacist, or wait in line. Without value alignment and positive reinforcement, the robot would simply grab the medicine and leave. By rewarding the robot for waiting patiently, Quixote corrects the behavior to align with human values.
Quixote is best suited for robots with a limited purpose that need to interact with humans to achieve it. The researchers describe it as a primitive first step toward general moral reasoning in AI. "We believe that AI has to be enculturated to adopt the values of a particular society, and in doing so, it will strive to avoid unacceptable behavior," Riedl said. "Giving robots the ability to read and understand our stories may be the most expedient means in the absence of a human user manual."
Editorial Notes