Homeostasis, Feedback Control, and Dynamic Causal Models

Graphical causal models (Spirtes et al., 2001; Pearl, 2009) provide a systematic framework for representing and testing causal hypotheses, and inspired the dominant philosophical account of causal explanation in the social and life sciences (Woodward, 2003). Yet, until recently, there was little work applying these models to dynamical systems, and this limits their relevance to a wide swath of modeling within science and engineering, including biology, cognitive science, and artificial intelligence. In this talk, I help bridge this gap by developing a dynamic causal model for the Watt governor, a textbook feedback-control system that inspired a paradigm shift in cognitive science, and which provides a starting point for thinking about processes such as homeostasis and reinforcement learning. While discussions of the governor typically emphasize its use of a negative feedback loop to keep a particular quantity fixed, the dynamic causal model reveals a much more illuminating story of how the governor is able to exploit this short-term feedback loop to ensure that a system's long-term energy supply adjusts to match varying demand. This has far reaching consequences for understanding how control feedback systems give rise to higher-scale intervention targets that are not available at lower scales. I conclude with some implications of the discussion for understanding homeostatic systems, advancing debates over mental representations, and clarifying the scope of the causal faithfulness condition