Expected Free Energy Model

Each policy $\pi$ (walk, skip, run, stroll) is scored by its expected free energy:

where the cost terms $C_k$ are Risk, Ambiguity, Energy, Social, Injury, and $|\text{Arousal}_\text{desired} - \text{Arousal}_\text{predicted}|$.

The policy with the lowest $G$ is selected. Costs (risk, energy, social penalty, injury, ambiguity, arousal mismatch) increase $G$; information gain decreases it, rewarding exploratory behaviour.

Policy Specifications

Each gait is parametrised by six numbers that feed into the EFE terms:

impact drives injury probability; signalAmp is proprioceptive richness (high for skip); conspicuous feeds the social penalty; complexity determines how much there is to learn.

The Child-Adult Crossover

Both a child and an adult share the same EFE equation and the same four policies. The difference is the weight vector $\mathbf{w}$. A child operates with high $w_i$ (curiosity) and low $w_s$ (social cost), so skipping wins: it is complex, proprioceptively rich, and novel. An adult raises $w_s$, $w_e$, and $w_j$, which penalises skipping enough that walking takes over. Try the Child and Adult presets and sweep mastery or normPressure to watch the crossover happen.

Notes

• This is a toy model. The policy specs are hand-tuned, not fitted to biomechanical or behavioural data.
• All context variables and weights are normalised to [0, 1] or [0, 2]. Absolute G values are arbitrary; only the ranking matters.
• The radar chart shows raw component values (before weighting). The waterfall shows weighted contributions to G.
• Sensitivity is measured by perturbing each context parameter by ±0.05 and checking whether the winner changes.