Utility extraction is the semi-automatic acquisition of decision maker's preferences about the different outcomes of a decision problem. Extracting human preferences would be of great importance in order to implement them in a Friendly AI, preventing AI’s goals differing from ours in case of a "hard takeoff". However, human values can be difficult to specify.
Research has focused on three different areas:
- eliciting the utility function based on a database of already elicited utility functions;
- iterative reﬁnement of the decision maker’s current utility function using a value of information approach;
- eliciting the utility function based on a database of observed behavioral patterns.
The last approach implies that preferences are reﬂected in the behavior, and that the decision maker is behavioral consistent. As real-world behaviors and decisions are often not consistent, methods based on such assumptions can extract only trivial utility functions.
Thomas D. Nielsen and Finn V. Jensen (Learning a decision maker’s utility function from (possibly) inconsistent behavior) were the first describing two algorithms that can take into account inconsistent behaviors. Inconsistent choices are interpreted as random deviations from an underlying “true” utility function.
Another possibility according to Luke Muehlhauser and Louie Helmis (The Singularity and Machine Ethics) is interpreting inconsistent choices as deviations produced by non-model-based valuation systems in the brain; information on when and to what extent model-based choices are “overruled” by the non-model-based valuation systems is provided by neuroscientific research.
Finally, another option is represented by "value learners", implemented agents flexible enough to be used even when a detailed specification of desired behavior is not known. As they can pursue any goal, they can be designed to treat human goals as final rather than instrumental goals (Learning What to Value). Agents are provided with a pool of possible utility functions and a probability distribution P given a particular interaction history: they can therefore calculate expected value over possible utility functions.
Further Reading & References
- A brief tutorial on preferences in AI by Luke Muehlhauser
- Learning a decision maker’s utility function from (possibly) inconsistent behavior by Thomas D. Nielsen and Finn V. Jensen
- The Singularity and Machine Ethics by Luke Muehlhauser and Louie Helmis
- Learning What to Value by Daniel Dewey