AGENTS with Theory of Mind for Intelligent Collaboration (ATOMIC)
Abstract
Agent architectures for understanding human behavior have gained considerable leverage by starting from a generative model capable of social reasoning and then reusing that model for inference, prediction, and intervention when interacting with people. This project will follow a similar methodology to create an agent-based system that combines a domain-independent model of collaboration with dynamic models of specific environments, tasks, and team members. We will implement a novel combination of existing AI frameworks: PsychSim, a decision-theoretic social simulation for Bayesian inference and optimization-based interventions, and Sigma, a cognitive architecture for unified learning, inference, and decision-making. Our agent will use these components recursively, giving it a Theory of Mind to reuse its own reasoning as a model of its human teammates. The representational language across all of these components will be in a declarative format that enables direct encoding of both general social-science theories and specific domain knowledge. This declarative language will also engender transparency of the agentÕs learned models and quantitative reasoning, transparency that is critical for successful human-machine team performance. The result will be an autonomous agent that reasons both bottom-up (from observations of teammates to inferences about their state and of the overall team) and top-down (from team-level goals and plans to predictions of teammatesÕ behavior and interventions to improve team performance). Reasoning in both directions will proceed from a first-principles understanding of collaboration that will allow the agent to identify messages that will improve the teamÕs performance even in the presence of perturbation or failure. The agentÕs expertise in both generating and understanding collaborative behavior, acquired through both knowledge engineering and data-driven methods, will all be in a declarative format that is both easily understood by other researchers and easily implemented within their own agent architectures. The encapsulation of social-science theory and learned knowledge within a unified framework will support empirical evaluation and quantification of the separate contributions of these various components across multiple dimensions and multiple domains of interest.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Dec 04, 2019
- Source ID
- W911NF2010011
Entities
People
- David V. Pynadath
Organizations
- Army Contracting Command
- Defense Advanced Research Projects Agency
- University of Southern California