nucleation

LIT-1181: Action and behavior: a free-energy formulation

3 min read

Properties5
StatusActive — read 2026-10-16 (NOTE-960) in full, from the publisher's open-access PDF; worth reading as the first full statement that action, like perception, minimizes a free-energy bound on sensory surprise, by fulfilling proprioceptive and visual predictions, with simulations of oculomotor stabilization and saccades, cued reaching, mountain-car and Lorenz trajectories. Read it knowing that the "ergodic arguments" of the abstract only equate entropy with time-averaged surprise, and that action is needed because the setup stipulates that action is the only way the agent changes the hidden states it senses; that action here pulls the sensed world toward a fixed model, sampling "what one expects to experience", and cancels unexpected perturbations so they are not perceived; and that the paper answers the worry about confirming a wrong model by appeal to existence ("agents who do not have the right priors cannot exist"). It has no mental or covert action.
Version2
Tagsneuroscience, probabilistic-modeling, embodied-cognition
SupportsCLAIM-339 — A judgement–observation loop cut off from action loses one route for evidence it did not select, and the free-energy principle's reason for minimizing surprise, that a persisting system must keep its states in bounds, needs action on the world, though a loop can confirm itself with action available too, CLAIM-341 — The loop that keeps a lucid agent's model in touch with what it models runs between judgement and observation: each judgement reshapes the observer whose later observations it then judges, and action on the world is not part of that loop, so active inference's action–perception cycle misplaces where the agent is active, CLAIM-342 — What keeps a judgement–observation loop from confirming itself is evidence whose arrival or scoring does not pass through the interpretation it tests; action is one source of such evidence, not a necessary one, and a loop that selects its own evidence can lock in with action available, CLAIM-338 — The selectivity active inference gives action, sampling what one expects to experience, is played in the coupled meaning–authority model by the authority space: influence decides which operators retrieve and so what the next judgement observes, which makes Friston's self-confirming sampling the model's authority-to-interpretation arrow and his agent a coupled loop run without independent scoring
Cited as “paper” byNOTE-960 — Friston et al. — action and behavior: a free-energy formulation

Graph View

Friston, K. J., Daunizeau, J., Kilner, J. & Kiebel, S. J. (2010). Biological Cybernetics 102(3):227–260. doi:10.1007/s00422-010-0364-z.

Key takeaways

From the reading, NOTE-960.

  • The argument. Existence requires low entropy; at equilibrium the ensemble density is the sojourn time, so sensory entropy is the long-run average of surprise (p. 230); free energy bounds surprise. Action must minimize it too because “the only way that the agent can change hidden states is through action” and agents “can only act by changing hidden states, not the sensory mapping per se” (p. 230). The abstract’s “mandated by ergodic arguments based on the very existence of adaptive agents” (p. 227) compresses these steps.
  • Action fulfils predictions. “there are no top-down motor commands; just prior expectations about proprioceptive input that are fulfilled at the spinal level” (p. 235). The generative model “has no notion of action” (p. 233).
  • Fixed model, re-sampled data. “The free-energy principle turns this optimization around and considers that the model (agent) is fixed and the aim is to minimize surprise by re-sampling the data” (p. 232); action amounts to “sampling selectively what one expects to experience” (p. 232).
  • Simulations. Oculomotor (36 photoreceptor channels; tight priors make action cancel a perturbing force so that “from the point of view of the agent, the target is not moving”, p. 239; a prior alone produces a saccade), cued reaching with optimized multisensory precision, mountain car learned with active inference “precluded” during training (p. 246), and Lorenz trajectories.
  • The anchor is existence. Confirming a wrong model is answered by “this weakness becomes the strength” (p. 232) and “agents who do not have the right priors cannot exist” (p. 256).
  • Scope. No mental or covert action; precision is inferred, and “may mediate attentional gain” (p. 248). All results are simulations.

Why this is here

The canonical early statement of the action–perception cycle the owner’s thesis rejects (CLAIM-341). The reading confirms the thesis’s description of it: all of the agent’s influence on what it senses runs through action, and the paper states the contrast in so many words, a fixed model with re-sampled data against the thesis’s changed observer.

For CLAIM-339 the quotation is exact and its use, that persistence needs action, matches the paper. But the paper’s action does not bring in evidence the loop did not select. It selects evidence to fit the model and cancels the unexpected. The mandate rests on a stipulation about the setup more than on the ergodic step. For CLAIM-342 the paper is a stronger case than NOTE-766: lock-in with action available is its design, and its external anchor is selection across agents.

Its oculomotor simulations model the action classic locked-in syndrome spares (CASE-107). They are not its only simulations.

Backlinks