Active Inference (Free Energy) Models

Description

Models users as agents minimising expected free energy, balancing goal-seeking (pragmatic value) with uncertainty-reducing exploration (epistemic value).

Active inference frames perception and action as minimising variational free energy, a bound on surprise under an internal generative model. Agents act to fulfil preferred observations and to reduce uncertainty, which naturally explains curiosity, information-seeking and exploration-exploitation trade-offs. Formally, models are partially observable Markov decision processes in which policies are scored by expected free energy, decomposed into epistemic and pragmatic value. For UX, it offers a single principled account of why users explore new interfaces, check status, seek confirmation and abandon when predictions fail. Applications remain experimental, but simulation studies can test whether a design reduces users' uncertainty where it matters.

Attributes

Loss type
Latent
Stage
Strategy, Design
Effort
Months

Steps

  1. Specify the user's generative model: hidden states, observations the interface provides, and actions available.
  2. Define preferences as prior beliefs about desired observations (e.g. task completed, no error message).
  3. Implement the model with a discrete-state active inference toolkit (e.g. SPM DEM routines or pymdp).
  4. Simulate behaviour under design variants, separating epistemic actions (checking, exploring) from pragmatic ones.
  5. Compare simulated with observed behaviour, for example frequency of status-checking actions.
  6. Use the model to identify where interfaces leave users uncertain and add observations that resolve it.

Sources

  • Friston, The free-energy principle: a unified brain theory? (2010), Nature Reviews Neuroscience
  • Parr, Pezzulo & Friston, Active Inference: The Free Energy Principle in Mind, Brain, and Behavior (2022)

Also known as

  • Free energy principle
  • Active inference
  • Predictive processing models

Categories

  • User Research
  • Cognitive Psychology & Perception