Action, uncertainty and policy selection
Active inference and expected free energy
Active inference extends perception-as-inference into action. An agent does not only update beliefs about the world; it chooses actions that are expected to reduce uncertainty and bring about preferred outcomes.
1. From free energy to action
In the previous page, free energy described how a belief can be updated to better explain sensory input. Active inference adds policy selection: the agent asks which action is expected to produce the best future evidence.
$\pi^\* = \arg\min_{\pi} G(\pi)$
Here, $\pi$ is a policy: a possible course of action. $G(\pi)$ is expected free energy: a score for how good or bad that policy is expected to be before the agent acts.
2. Interactive policy selection
The agent can move left, right, up or down. Some locations are informative because they reveal hidden state information. One location is preferred because it contains the goal. The selected action is the one with the lowest expected free energy.
3. Pragmatic and epistemic value
A useful teaching decomposition of expected free energy is:
$G(\pi) \approx \text{expected cost} - \text{expected information gain}$
Policies are attractive when they are expected to reach preferred outcomes and reduce uncertainty. This gives active inference its characteristic balance between exploitation and exploration.
Pragmatic value
Will this action lead to preferred outcomes?
Epistemic value
Will this action reduce uncertainty?
Expected free energy
A policy score combining future cost and information gain.
Policy precision
How strongly the agent commits to the best-scoring policy.
4. Perception, action and the world
Active inference closes the loop. The agent has beliefs about hidden states, uses those beliefs to predict future outcomes, selects actions, samples the world and then updates its beliefs again.
hidden state estimates
possible actions
new evidence
Why this matters for our lab
Active inference provides a computational language for embodied intelligence, adaptive behaviour and psychiatric symptoms. In our work, the same principles can be used to think about perception, decision-making, motor control and model fitting. For example, a Pong agent can select paddle movements by minimising expected free energy, while a clinical model can ask how altered precision or synaptic gain changes the way beliefs and actions are updated.