Active inference and artificial agents
Explanation
Active inference is the theoretical framework developed mainly by Karl Friston, of University College London, since the 2000s. Although it originated as a neuroscientific theory (to understand the brain), it has extended as a general principle of agency applicable both to biological systems and to artificial agents. Its mathematical basis is the Free Energy Principle, which Friston proposes as a universal organising principle of self-organising systems.
The central idea: every self-organising system must keep its states within a limited range (homeostasis, viability). This requires minimising a mathematical quantity called variational free energy, which is an upper bound of surprise (the improbability of observations given the agent's internal model). Minimising free energy is equivalent to improving the internal model of the world so that observations are less surprising.
There are two ways to minimise free energy: (1) perception: updating the internal model so it explains observations better (traditional perceptual inference); (2) action: changing the world so it matches the model's predictions (active inference proper). Thus, perception and action are two sides of the same process of minimising free energy. The brain/agent is a machine that actively seeks confirmations of its models through directed action.
This framework explains many things: why the brain makes predictions (to minimise surprise), why we explore (reducing uncertainty improves the model), why we have curiosity, why we avoid catastrophic states, why attention goes where predictions are most uncertain. It can be formulated at various levels: quantum-thermodynamic, biophysical, neuronal, cognitive, social.
In AI, active inference offers a framework for designing agents with multiple advantages over traditional reinforcement learning: it combines perception, action and learning in a unified principle; it promotes intrinsic exploration (curiosity, uncertainty reduction) without needing elaborate external rewards; it favours robustness and generalisation; it allows prospective planning (imagining consequences of actions to minimise expected future free energy). Initiatives such as the VERSES Research Lab work on active-inference-based AI.
For the theory of consciousness, active inference has interesting implications. Friston has connected the framework with questions of self and consciousness: the self would be a hierarchical model that minimises free energy including predictions about one's own body and interactions; consciousness could be associated with certain high hierarchical levels that maintain global precision weighting over the system. Authors such as Anil Seth, Thomas Metzinger have articulated theories of consciousness based on active inference. Although the framework does not solve the hard problem on its own, it offers a mathematically rigorous vocabulary for many questions about perception, action, self-control, and promises to be one of the bases for a unified science of biological and artificial minds. Its influence in coming years will probably be growing.
Strengths
- Unified mathematical framework for perception, action and learning.
- Integrates exploration and exploitation in a single principle.
- Superior interpretability to classical RL.
- Natural connection with neuroscience, phenomenology and enactivism.
- Applicable to robotics and embodiment-first AI.
Main critiques
- High computational costs in full implementations.
- Broad version of the FEP accused of being unfalsifiable.
- Phenomenal interpretation of the self as a model still open.
- Practical distance from mainstream deep-learning frameworks.