Participation Learning

In DemocracySim, agents do not learn strategic ballot manipulation. They learn whether to participate in later elections.

Position In The Election Loop

Participation learning happens after an election outcome has been translated into fees, rewards, and per-agent relative asset changes. The voting rule therefore affects participation only indirectly: through the outcomes it produces and the incentives that follow from them.

For the election pipeline around this update, see core_mechanics.md.

Decision Policy

Each agent carries a participation state q_participation. Participation probability is logistic in that state:

• p = sigmoid(participation_beta * q_participation)

The actual decision is then sampled from a Bernoulli draw with probability p.

Current Default Signal Mode

The default mode used in the thesis-oriented configuration is:

• participation_signal_mode = group_relative_delta_rel_party

Its logic is group-relative rather than purely individual. For each preference group g, the model computes the mean relative asset change of the eligible agents in that group:

• mu_g = mean(delta_rel of eligible agents in group g)

These group means are centered against the mean across groups:

• mu_groups = mean(mu_g across groups)
• group_component_g = (mu_g - mu_groups) * n_g / (n_g + participation_signal_group_shrink_k)

For an individual agent i, the signal then combines that group-relative term with fee salience:

• fee_component_i = - participation_signal_fee_weight * fee_rel_i for participants, else 0
• signal_i = clip(group_component_g + fee_component_i, ±participation_signal_clip)
• q_push_i = signal_i

The update is:

• q_participation <- q_participation + participation_alpha * q_push_i

If configured, q_participation is clipped symmetrically by participation_q_max.

Why This Signal Is Structured This Way

The model includes a costly participation decision and collective outcome effects. Rewards and penalties are shared within groups, while the participation fee is paid only by participants. That creates a free-rider structure.

The default signal mode therefore does not treat participation learning as pure individual reward chasing. It frames the update as a combination of:

• how an agent's preference group performed relative to other groups
• how much participation cost the agent personally had

Main Knobs

The main parameters for participation learning are:

• participation_alpha
• participation_beta
• participation_init_q
• participation_q_max
• participation_signal_mode
• participation_signal_fee_weight
• participation_signal_group_shrink_k
• participation_signal_clip
• participation_baseline_alpha

Implementation Reference

The group-level signal construction lives with the election logic in Area. Per-agent q-state updates live in VoteAgent, and the participation decision policy is exposed in Strategies.

Compatibility Notes

Other signal modes still exist in code for compatibility and comparative experiments, including raw_delta_rel and group_centered_delta_rel_plus_fee. The main current path is group_relative_delta_rel_party.