QUANTIFIABLE EXPECTABILITY AND MEASURABLE INTENTIONS- COMPUTATIONAL CALIBRATED TRUST IN HUMAN-AI JOINT ACTIVITY
Abstract
Human-AI calibrated trust needs quantifiable short and long term computable expectability. We define expectability as being anticipatory, understandable, and precautionary with regards to one entity’s behavior by another entity necessary for joint action. We posit that without expectability, trust cannot be established let alone be maintained in any partnership. Further, expectability must be founded upon intentions – the what, why, how, and when governing our behaviors, decisions, and actions. Calibrated trust requires inferring of intent of and by each partner to ascertain how close or far the inferred intent is from the actual intent and hence trust. Inferred intentions in this effort involves learning the rewards and reward structures employed by the target entity as part of shared context. Deviations from the expected rewards and structures from the actual forms the computational basis for trust calibration. To date, there has been no comprehensive solution that provide quantifiable, measurable mechanisms with which to formulate computable expectability and ultimately, any well-founded approach to calibrated trust. Our goal is to define, develop, and demonstrate measurable intentions for computational calibrated trust in human-AI partnerships applied to joint activity and-or shared context for the domains of intelligence and digital data analysis and tactical and strategic decision making in computer games and simulations. Our solution is founded upon (a) new concepts and algorithms for individual (human or AI) and teambased multi-agent inverse reinforcement learning (IRL), (b) a novel Preferential Trajectory IRL (PT-IRL) that distinguishes between multiple different individuals and teams by their behavior overcoming existing linearity relationship assumptions in IRL, and (c) a formal definition for quantifiable interference and intentionality gap that reveals and explains how each team member is adapting and learning during joint activity and decision-making.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 07, 2023
- Source ID
- FA95502210022
Entities
People
- Eugene Santos
Organizations
- Air Force Office of Scientific Research
- Board of Trustees of Dartmouth College
- United States Air Force