Active inference with generalised policies. This example shows how paradoxical but adaptive behaviour (moving away from a target to secure it later) emerges from simple priors on the motion of hidden states. These priors are encoded in a cost function (upper left). The form of the agent's (generalised) policy ensures that divergence is positive or friction is negative in regions of positive cost, such that the car expects to go faster. The inferred hidden states (upper right: position in blue, velocity in green, and friction in red) show that the car explores its landscape until it encounters the target and friction increases dramatically to prevent it escaping (i.e., falling down the hill). The ensuing trajectory is shown in blue (lower left). The paler lines provide exemplar trajectories from other trials, with different starting positions. In the real world, friction is constant (one eighth). However, the car expects friction to change with position, enforcing exploration or exploitation. These expectations are fulfilled by action (lower right), which tries to minimise free energy.
