Information-theoretic decomposition of embodied and situated systems

Area of Science:

• Cognitive Science
• Robotics
• Artificial Intelligence
• Information Theory

Background:

• Embodied cognition emphasizes real-time, nonlinear bodily interaction with the environment for knowledge development.
• Artificial evolution and neural networks are used to study complex adaptive systems.
• Information-theoretic measures quantify dependencies and information flow within systems.

Purpose of the Study:

• To investigate nonlinear interactions in an agent-environment system using robots learning a wall-following task.
• To analyze the functional neural structure and behavioral dynamics of evolved robots.
• To apply information-theoretic measures for a detailed decomposition of system complexity.

Main Methods:

• Populations of robots with artificial neural networks underwent artificial evolution for a wall-following task.
• Time series data from perceptual and motor neurons were recorded from selected evolved robots.
• Mutual information and transfer entropy, including their local forms, were used to analyze neural and behavioral data.

Main Results:

• Different information-theoretic measures revealed complementary insights into robot-environment interactions.
• The analysis uncovered nonlinear dependencies and information flow within the agent-environment system.
• The study identified characteristics of the robots' functional neural structure and underlying dynamics.

Conclusions:

• Information-theoretic measures provide a powerful framework for decomposing complex agent-environment systems.
• These measures effectively capture the intricate nonlinear relationships governing robot behavior and neural dynamics.
• The findings support the embodied and situated view of cognition by demonstrating environmental interaction's role in structuring neural systems.