Socio-cognitive Architectures for Adaptable Autonomous-Systems

What Is It?

Socio-cognitive architectures form the basis for software platforms to be installed on adaptable autonomous systems. Socio-cognitive architecture focuses on cognition about other agents, including their beliefs, desires, intentions, and obligations. It hinges on computationally instantiated theory of the entire cognitive system, including perception, cognition, learning and action, constrained by computational, neural, and psychological findings.

How Does It Work?

Socio-cognitive architectures reason directly about the mental states of other agents—either human or non-human—by maintaining “mental models” corresponding to the mental perspective of those agents.

What Will It Accomplish?

Basic research will support autonomous systems that are capable of enhanced coordination with both human and machine teammates; intelligent systems that are capable of natural language interactions through dialogue; systems that are capable of rapid learning via discernment of teacher’s intent; and systems that teach more effectively by understanding student intent and comparing against skilled performance.

The Office of Naval Research (ONR) has prioritized its investment in cognitive architectures for more than 20 years. Among its many success stories, cognitive architectures have been broadly applied to training and education in the development of intelligent tutoring technologies used at Department of Defense-dependent schools. Additionally, validated cognitive models of naval operator performance are currently being applied to human-to-computer interface issues as a cost-effective technique for exploring new concepts in naval displays.

Socio-cognitive architectures are outfitted with mechanisms corresponding to neural and psychological theories regarding the human ability to reason about the mental states (e.g., beliefs, desires, intentions, etc.) of other humans. Predicting and explaning the behavior of others in terms of their mental states is sometimes called “mind reading.” In general, mind readers must be able to construct and maintain two “mental models”—one of the world as it is seen first-person, and one as it is seen from the perspective of the target agent to be predicted.

The types of mental operations that can be performed within or between these mental models remains an open issue for the psychology community, but the computational cognitive modeling community is currently moving to extend or adapt current cognitive architectures to be able to hold two or more such models in memory.

Research Challenges and Opportunities:

  • Representation of mental states within cognitive architecture: Mental simulation of other agents, bodies of conceptual knowledge about other agents, hybrid techniques, integrating learning with mental-state reasoning for rapid behavior prediction and explanation
  • Integrating probabilistic and relational knowledge: Real-time discernment of intentional action representing uncertainty about the mental states of others, and disambiguation of spoken utterances in natural language dialogue
  • Socially mediated learning: Self-reflection/diagnosis on task performance and mental simulation of alternatives in future performance; discernment of teacher intent during learning-by-imitation to facilitate quicker knowledge/skill acquisition; and teacher discernment of student intent during performance to provide effective remediation

Paul Bello
(703) 696-4318
paul.bello@navy.mill

 

* Some pages on this website provide links which require Adobe Reader to view.