Implicit Communication in Human-Machine Collaboration
Abstract
Pragmatics is an area of linguistics founded on the idea that what we say has implicit meaning. The central tenet of this work is that it is not just what we say that has implicit meaning: what we do has implicit meaning as well. We infer someone’s intentions, reason about their future actions, build a model of what they are capable of doing, and even infer their emotional state, all based on what they are doing. Our goal is to enable machines to take part in this implicit communication that happens so naturally among humans. Our thesis is that a machine’s ability to infer implicit meaning in a collaborator’s behavior, as well as to purposefully choose actions that express a certain implicit meaning to a collaborator, are both crucial steps in enabling better human-machine collaboration. We introduce a Bayesian formalism that relates expression and inference, and support it through preliminary examples in inferring and expressing intent (goals) from ongoing motion. To develop generic tools for inference and expression, we will look beyond goals, to three thrusts: strategies or constraints, capability, and emotional state. We chose these thrusts because of their diversity with respect to how concretely they can be captured through mathematics, but also because they are highly practical. They enable advances in shared workspace interactions with machines as well as remote operation of vehicles via shared autonomy – machines will be able to adapt to strategies, infer and maintain desired constraints, or customize their assistance to the human’s fatigue or frustration level. They also enable advances in interactions with disembodied agents, e.g. helping humans understand what kinds of task they can support. In the midst of amplifying worries about the future of AI, we see such tools that support transparency of internal state as a critical next step.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 11, 2017
- Source ID
- FA95501710308
Entities
People
- Anca Dragan
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of California Regents