ICI Local Effects Elicitor
Overview
The ICI Local Effects Elicitor is a quantitative knowledge‑elicitation capability provided by Hellixia (the LLM‑based assistant in BayesiaLab). Its purpose is to assess how each parent variable individually influences a child node, following the Independence of Causal Influence (ICI) framework.
Instead of asking the expert or the LLM for an entire joint Conditional Probability Table (which grows multiplicatively with the number of parent states), the elicitor focuses on one parent at a time. For each parent, it asks a Sole‑Active‑Cause Impact Question:
“What impact does Cause i have on the Overall Effect when it is the sole active cause, or when all other causes are in their Neutral State?”
The answer produces that parent’s Local Effect, a distribution over the effect scale (e.g., Low / Normal / High, or a numeric range). All Local Effects are then merged by the chosen Combination Function (OR, Sum, Vote, Average, etc.) into an Overall Effect, which finally maps to the child’s states.
Why this is efficient
- The number of probabilities to elicit grows linearly with the number of parents rather than multiplicatively.
- Each parent’s Local Effect can be estimated independently, making the task cognitively simpler for both human experts and LLMs.
- The elicitor pairs with Hellixia Prior Elicitation (which estimates marginal priors) to build a full, consistent CPT.
Neutral States
For the Sole‑Active‑Cause question to be well‑posed, each parent should have a Neutral State – a state whose Local Effect is the identity element of the Combination Function (e.g., “no impact” for Sum, “false” for OR). When a parent is neutral, it contributes nothing to the Overall Effect, isolating the influence of the active cause.
Workflow Example
If data are later available, Learning Hidden Local Effects with EM can refine the Local‑Effect distributions that were initially elicited, bridging expert knowledge with data‑driven learning.