Neuro-Inspired Active Inference Agent

Modular, general-purpose generative architectures for intelligent behaviour.

Overview

This page outlines the architecture and inference procedures used in our neuro-inspired active inference agent, developed as part of our work on general-purpose intelligent systems. The agent is based on predictive coding, variational message passing, and the free energy principle.

Generative Model Structure

The agent is structured as a 7-module generative model, with components handling perception, dynamics, action selection, and learning. The latent state evolves according to:

\[ x_{t+1} = f(x_t, a_t) + \omega_t, \quad y_t = g(x_t) + \epsilon_t \]

where \( x_t \) is the hidden state, \( a_t \) the action, \( y_t \) the sensory input, and \( \omega_t, \epsilon_t \) are noise terms.

Free Energy Minimisation

The agent selects actions and updates beliefs to minimise variational free energy:

\[ \mathcal{F}(q) = \mathbb{E}_q [\log q(x) - \log p(x, y)] \]

Action is selected by planning over expected free energy (EFE):

\[ G(a) = \mathbb{E}_q [ D_{KL}[q(s) \| p(s)] - \mathbb{E}_{q(s)}[\log p(o|s)] ] \]

Inference Loop

Each timestep consists of:

Receive sensory input \( y_t \) (possibly partial/noisy)
Infer hidden state \( x_t \) via variational update (e.g., Laplace)
For each candidate action \( a \), roll out future trajectory and compute expected free energy \( G(a) \)
Select action that minimises \( G(a) \)
Apply action \( a_t \), observe outcome, repeat

Pong Agent

The Pong agent plays a 2D paddle game using an active inference loop. It observes a partial and noisy visual input of the ball and paddle positions, and uses belief updates and action selection to intercept the ball.

Observation: Noisy 1D ball and paddle coordinates
Hidden state: Belief over ball trajectory and paddle location
Action: Move paddle up/down
Planning: Multi-step EFE rollouts to simulate future ball-paddle outcomes
Control: Select actions that bring ball contact into high-probability state space

The agent uses a low-rank Laplace posterior over state variables and updates belief online during gameplay.

Drone Agent

This agent navigates a 2D/3D environment (gridworld or room maze) toward a target using visual or proprioceptive cues. The belief update and action selection loop is as follows:

Observation: egocentric image or sparse distance sensors (noisy)
Hidden state: agent's location, orientation, and velocity
Actions: move forward, turn left/right, hover
Planning: expected free energy minimisation over predicted future states using dynamics model
Goal: minimise uncertainty and reach preferred state (goal location)

Inference and control are implemented with online updates to the posterior over latent states and forward rollout over dynamics.

Resources

For the dynamical-systems background, our interactive Helmholtz–Hodge and NESS lab explores gradient and solenoidal flow, stationary probability currents, and conservative versus dissipative dynamics. It also explains the connection to stationary surprisal and the assumptions behind that decomposition.