Metacognition for Unknown Situations and Environments (MUSE)

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Robotics

Background:

• Human adaptability relies on metacognition (awareness and regulation of cognition).
• Current autonomous agents lack adaptability in novel environments.
• Metacognition is hypothesized as key for autonomous agent flexibility.

Purpose of the Study:

• To investigate metacognition's role in autonomous agent adaptation.
• To develop a framework integrating metacognitive processes into agents.
• To enhance agents' competence awareness and strategy selection.

Main Methods:

• Proposed the Metacognition for Unknown Situations and Environments (MUSE) framework.
• Implemented MUSE using world modeling and large language models (LLMs).
• Agents self-assess competence and use it to guide strategy selection.

Main Results:

• MUSE agents show high competence awareness and improved self-regulation.
• Agents effectively solve novel, out-of-distribution tasks.
• MUSE outperforms model-based reinforcement learning and prompt-based LLM agents.

Conclusions:

• Metacognition significantly improves autonomous agent adaptability.
• MUSE framework enables agents to tackle unfamiliar challenges.
• Cognitive and neural system inspiration offers a path beyond data-intensive models.