Toward Cognitive Realism in Game Theoretic Models of Social Behavior
Abstract
The overall objective of this project is to develop methods that will allow an artificial agent to use representations of people, places or thing that vary in their level of abstraction in relation to the agent's psychological distance. Construal Level Theory (CLT) posits that as psychological distance decreases representations of people, plans, events, and things become more concrete and well-defined. This project uses ideas from CLT to create a simulated agent capable of solving complex social and resource problems. Specifically, this research will 1) Develop a computational process for creating construals of varying abstraction; 2) Extend the construal process to learning construals representing categories of simulated people; 3) Integrate CLT with Bayesian games to generate agent social strategies; and 4) Perform simulation testing and metric development in a rich, complex environment. Variations of agglomerative clustering will be used to create abstraction hierarchies of items and people in simulation. Individual abstractions will act as high or low-level construals. The features composing these construals will influence the agent's planning. Construals will be used to populate a game structure by acting as a prior over the distribution of the players types and/or reward outcomes. This model will be used to create a strategy for optimizing interactions with the other agents. The developed computational process may inform hierarchical planning and reinforcement learning, increasing the effectiveness and trust in a variety of AI related applications.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 21, 2022
- Source ID
- FA95502110197XX0
Entities
People
- Alan Wagner
Organizations
- Air Force Office of Scientific Research
- Pennsylvania State University
- United States Air Force