Figure 2.

Decision information heat map in a 5 × 5 gridworld with $\mathcal{A}\hspace{0.17em}:\hspace{0.17em}\{\uparrow ,\leftarrow ,\downarrow ,\to \}$. Annotated values show ${\mathrm{\Im }}_D^{{\pi }_{\mathcal{F}}}(s)$ in bits. The policy presented is optimal with respect to free energy, i.e ${\pi }_{\mathcal{F}}={\arg \hspace{0.17em}\min \hspace{0.17em}}_{\pi }{\mathcal{F}}^{\pi }(s;\beta )$. The arrow lengths are proportional to the conditional probability π(a|s) in the indicated direction, for convenience the prior $\widehat{\hspace{0.17em}p}(a)$ is shown instead of the policy in the yellow goal state. Reward-maximizing behaviour (β = 100): in (a) the goal is in the corner and in (b) the goal is on the diagonal between the corner and the middle. Behaviour where information processing is constrained (β = 0.1): (c) the goal is in the corner and in (d) it is diagonally adjacent to the corner.