Concepts

This page explains the conceptual model used by PolicyFlux.

Mental model

PolicyFlux simulates repeated policy decisions in a structured institutional setting:

1. actors evaluate bill positions,
2. influence layers transform baseline preferences,
3. transformed preferences are aggregated,
4. a voting strategy produces outcomes,
5. engines collect aggregate statistics.

This separation lets you test assumptions in isolation.

Policy space

Actors and bills are represented in the same n-dimensional policy space.

• actors: preferences/positions,
• bills: policy positions,
• voting behavior: affected by distance, utility, and institutional influences.

In practice:

• lower actor-bill distance usually means higher support probability,
• higher policy_dim increases representational richness,
• larger spaces often require more actors/iterations for stable comparisons.

Conceptually, preference compatibility can be thought of as:

$$ ext{support tendency} \propto -\text{distance}(\text{actor position}, \text{bill position}) $$

Layers

Layers are composable influences on vote behavior:

• ideal point,
• public opinion,
• lobbying,
• media pressure,
• party discipline,
• government agenda.

Each layer encodes a distinct mechanism. For example:

• public opinion shifts support in line with electorate pressure,
• lobbying introduces targeted directional pressure,
• party discipline constrains individual deviation,
• agenda control modifies which proposals reach final voting.

Use layer toggles to move from minimal baselines to richer institutional dynamics.

Aggregation

Layer outputs are combined with an aggregation strategy:

• sequential,
• average,
• weighted,
• multiplicative.

Choose strategy based on your analytical objective:

• average: balanced and interpretable baseline,
• weighted: explicit relative importance across layers,
• sequential: ordered, path-dependent composition,
• multiplicative: interaction-sensitive attenuation/amplification.

Keep aggregation fixed while testing single-layer effects to preserve interpretability.

Engines

Simulation engines run repeated sessions and expose outcome metrics:

• deterministic,
• sequential Monte Carlo,
• parallel Monte Carlo.

Deterministic runs are ideal for debugging and baseline checks. Monte Carlo runs are better for estimating uncertainty and variance.

Outcomes and interpretation

Primary outputs typically include:

• pass_rate,
• accepted and rejected counts,
• session-level summaries (backend dependent).

Interpret results comparatively:

• compare scenarios with shared seeds and scale,
• alter one variable group at a time,
• inspect both central tendency and spread.

Concept-to-configuration map

Concept	Typical controls
policy-space scale	num_actors, policy_dim
reproducibility	seed
institutional assumptions	preset factories, AdvancedActorsConfig
influence composition	LayerConfig include flags, strengths, layer_names
combination rule	aggregation_strategy, aggregation_weights
runtime method	deterministic vs Monte Carlo engine backends

Continue with Configuration for parameter-level setup patterns.