An Empirical Characterization of Parsimonious Intention Inference for Cognitive-level Imitation Learning

Imitation learning is a promising route to better collaboration between humans and artificial agents. It will be most effective if the agent has some cognitive-level “understanding” of a human demonstrator’s intentions. Inferring intent is an example of abductive reasoning, wherein an agent explains the available evidence based on causal knowledge. Good explanations should satisfy some notion of parsimony (“Occam’s razor”), but the optimal notion of parsimony is often application-specific. We compare several such notions in the context of intention inference, using a robotic imitation learning scenario and the Monroe County Corpus, a standard benchmark in intention inference. Our results suggest that the most popular notions of parsimony in general are not necessarily appropriate for intention inference in particular.