Active Inference and the Physics of Agency
A Unified Theoretical Approach to Predictive Control and Informational Complexity
The Thermodynamic Laws of Agency provide a foundational perspective on agency, explicitly connecting it to physical and informational principles. Integrating these laws with Karl Friston's Active Inference framework, particularly the Variational Free-Energy Principle (FEP), offers a robust, unified model of agency grounded in predictive control, thermodynamic constraints, and informational complexity. This synthesis clarifies the mechanisms by which agents interact dynamically with their environments, continuously adapting to sustain their existence and functionality.
Law 1: Control Work
Agency Law: Exercising intentional control over outcomes requires physical work proportional to kybits exerted.
Active Inference Integration: Under Friston's Active Inference framework, agents minimize variational free energy through actions specifically aimed at reducing prediction errors between their internal predictive models and actual sensory data. The informational complexity captured by kybits directly corresponds to the precision-weighted prediction errors that agents systematically attempt to minimize. The greater the complexity and precision needed to align predictions with sensory inputs, the more substantial the physical and informational work required, thereby establishing a direct link between agency and the energetic costs of predictive accuracy.
Law 2: Agency Decay
Agency Law: In a closed system without external energy input, agency inevitably diminishes.
Active Inference Integration: Within Friston's framework, maintaining agency requires ongoing informational and energetic exchanges with the environment to continuously minimize variational free energy. When isolated, agents experience an accumulation of prediction errors as their internal predictive models become less accurate due to an absence of external inputs necessary for updating and refining those models. Consequently, the internal structures supporting predictive capacity degrade progressively, reducing the agent’s ability to effectively control or influence its environment, thus illustrating the inevitable decay of agency without continuous external resources.
Law 3: Agency Limits
Agency Law: Perfect frictionless control is physically impossible.
Active Inference Integration: Friston explicitly acknowledges that minimizing variational free-energy to a perfect degree is impossible due to inherent uncertainty, environmental fluctuations, noise, and model inaccuracies. The variational free-energy minimization process is inherently approximate and bounded by the unavoidable mismatch between internal predictions and external realities. Thus, even the most finely tuned and energetically robust predictive systems retain residual errors, making absolute predictive precision and frictionless control fundamentally unattainable.
Unified Perspective
This explicit integration underscores that agency fundamentally revolves around the continuous minimization of variational free energy through predictive control mechanisms, inherently limited by informational complexity and thermodynamic realities. The unified theoretical perspective generated by this integration explicitly accounts for physical constraints, informational capacities, computational resources, and inherent uncertainties. Such a perspective not only clarifies the inherent limits and dynamic nature of agency but also provides practical insights into developing realistic and adaptive models of intelligent systems capable of robust interaction and survival in complex environments.
References
Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience, 11(2), 127-138.
Friston, K. J., Parr, T., & de Vries, B. (2017). The graphical brain: Belief propagation and active inference. Network Neuroscience, 1(4), 381-414.
Friston, K. (2013). Life as we know it. Journal of the Royal Society Interface, 10(86), 20130475.
Parr, T., & Friston, K. J. (2019). Generalised free energy and active inference. Biological Cybernetics, 113(5-6), 495-513.
Friston, K. J., Da Costa, L., & Parr, T. (2020). Some interesting observations on the free-energy principle. Entropy, 22(12), 1387.
Clark, A. (2015). Surfing uncertainty: Prediction, action, and the embodied mind. Oxford University Press.