Reverse engineering the centered self

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Decision Theory

Background:

• Intelligent agents often solve problems from a self-centered perspective, a process not explicitly modeled by standard frameworks like partially observable Markov decision processes (POMDPs).
• Existing models do not fully capture the agent's subjective experience of centering itself within its environment to define and solve problems.

Purpose of the Study:

• To develop a computational framework, termed ePOMDPs, that explicitly models agent-centered problem-solving.
• To extend this framework with a meta-ePOMDP agent within a hierarchical Bayesian approach to model how agents select problems.
• To provide a computational basis for understanding how agents form a sense of self in problem-solving contexts.

Main Methods:

• Extended the standard POMDP framework to create the ePOMDP formalism, representing agent-centered planning and action.
• Developed a meta-ePOMDP agent using a hierarchical Bayesian framework to model problem selection.
• Implemented and tested meta-ePOMDP agents on specialized 'centering game' tasks.

Main Results:

• The proposed ePOMDP models successfully explain key signatures of agent-centered problem-solving.
• These models account for difficulties in navigating problem representation spaces, a challenge not addressed by alternative models.
• The computational framework demonstrates efficacy in capturing agent-specific perspectives in decision-making.

Conclusions:

• The ePOMDP and meta-ePOMDP frameworks offer a novel approach to modeling agent-centered cognition.
• This work provides a computational foundation for understanding the emergence of self-representation in intelligent agents.
• The findings suggest a pathway toward developing more sophisticated AI systems capable of subjective experience and self-awareness.