Quals research talk: Planning with and inferring moral and social norms with temporal logic
Abstract
Robots and other artificial agents are increasingly being considered
for domains requiring complex decision-making and interactions with
humans. These artificial agents will need the ability obey human moral
and social norms, even though these norms often conflict. These
agents will also need the ability to learn such norms, both by
instruction (in natural language) and by observing the behaviors of
other agents.
Inverse reinforcement learning (IRL) (observing agent behaviors and
inferring a reward function that ``explains'' those behaviors) is
often touted as a solution to the norm learning problem. We argue
that IRL is inadequate for the task, since it is (1) incapable of
learning temporally complex norms, and (2) not easily interpretable.
To address these problems, we substitute for the reward function a set
of statements in linear temporal logic (LTL). We propose algorithms
for maximally satisfying a set of LTL norms (even when they conflict).
We also propose an approach for inferring LTL norms from observed
behavior, analogous to IRL.