Recognizing the Goals of Uninspectable Agents
Abstract
Effective interaction between agents requires reasoning about other agents internal states. In some situations, such as in the case of multiagent systems with a shared policy, agents may have full knowledge of each other's knowledge, preferences, and goals. When interacting with humans or independent artificial agents, however, such direct inspection is not available. Instead, agents must model the internal states of their compatriots through observations of their behaviors in the world. In humans, such reasoning is called theory of mind (ToM). It has been argued that ToM reasoning can improve performance for artificial agents in team scenarios, as well. Here, we compare the performance of a model of ToM (Analogical Theory of Mind; Rabkina et al., 2017) with that of a state-of-the-art goal recognition system (Holler et al., 2018) on goal recognition tasks of increasingly uninspectable agents. We show that ToM reasoning is beneficial for agents when direct access to the internal states of their compatriots is not available.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 01, 2020
- Accession Number
- AD1126419
Entities
People
- Irina Rabkina
- Jason Wilson
- Ken Forbus
- Laura M Hiatt
- Mark Roberts
- Pavan Kathnaraju
Organizations
- Drexel University
- Northwestern University
- United States Naval Research Laboratory