>[!warning] >This content has not been peer reviewed. # Agents minimize the difference between model and sensation **Free Energy Principle (Friston 2010 and later):** Organisms (and more generally, systems that maintain themselves away from equilibrium) act so as to minimise **free energy** — a quantity that bounds the difference between their internal model of the world and the sensory input they receive. In practice this means: **predict** the incoming signal, and minimise prediction error (or its long-term average). So “life” is a process that actively keeps its internal state consistent with the environment by reducing surprise (Friston 2010; Friston & Kiebel 2009). --- ## What it is - **Free energy** $F$: In variational inference, $F \geq -\ln P(\text{data})$ (surprise). Minimising $F$ is equivalent to maximising the evidence (model fit) or minimising prediction error. - **Active inference:** Agents don’t just infer; they **act** to change their sensory input so that it matches their predictions. So behaviour is driven by the same principle: keep prediction error low. - **Implication:** “Intentionality” and “goal-directed behaviour” can be reframed as **resource-allocation strategies**: the agent invests energy to make the world (or its sampling of it) match its model, thereby reducing the ongoing “cost” of prediction error. --- ## How RRT/RST uses it - **Biological Substrate Triangle:** Life is a **cluster of relations** that actively minimises its **local noise floor** $N$ to increase its **[[Fidelity]]** $\mu$. The **[[Resource Triangle]]** $W^n = \Omega^n + N^n$ applies: the organism has a finite budget $W$; by predicting and acting, it allocates so that more of the budget goes to “signal” $\Omega$ (expected, explained) and less to “noise” $N$ (unexpected, costly). So **[[Fidelity]]** $\mu = \Omega/W$ rises when the agent minimises prediction error. - **Intentionality from the ledger:** “Life” and “intelligence” are **highly advanced resource-allocation strategies** ([[The minimum cost to erase a bit]]): they lower the effective Landauer bill by **predicting** the noise before it hits — i.e. by making the world more compressible (simpler model) and the sensory stream more predictable. So the Free Energy Principle is the **biological implementation** of the same optimisation that, at the physical level, favours compressed states (Solomonoff) and lawful structure (RST). --- ## Links | Concept | Note | |:---|:---| | Budget, signal vs noise | **[[Resource Triangle]]**, **[[Fidelity]]** | | Cost per bit | **[[The minimum cost to erase a bit]]** | | Substrate, format | **[[Format]]** | --- ## References - Friston, K. (2010). *The free-energy principle: a unified brain theory?* Nat. Rev. Neurosci. **11**, 127–138. [DOI](https://doi.org/10.1038/nrn2787) - Friston, K. & Kiebel, S. (2009). *Predictive coding under the free-energy principle.* Philos. Trans. R. Soc. Lond. B **364**, 1211–1221. [DOI](https://doi.org/10.1098/rstb.2008.0300) - Friston, K., Kilner, J. & Harrison, L. (2006). *A free energy principle for the brain.* J. Physiol. Paris **100**, 70–87. [DOI](https://doi.org/10.1016/j.jphysparis.2006.10.001) - Clark, A. (2013). *Whatever next? Predictive brains, situated agents, and the future of cognitive science.* Behav. Brain Sci. **36**, 181–204. [DOI](https://doi.org/10.1017/S0140525X12000477) (Review: predictive processing, free energy.)