Neuro‑symbolic systems must select answers from multiple expert agents or knowledge sources that are truly relevant to the current query. As the knowledge network expands, using only relevant experts becomes essential for accuracy and scalability. We adopt the legal theory of evidence to define relevance: an answer is relevant only if it is both probative and material for the matter at hand.\
Each expert is represented by a feedback fuzzy cognitive map (FCM), a directed causal graph that can be mixed into a common knowledge network. The mixing first computes each expert’s equilibrium or limit‑cycle prediction for the given input, then assesses probative status by measuring differences or variations among the answers.\
A Bayes‑like theorem is introduced to de‑mix the combined FCM into posterior‑like expert FCMs:
$$ P(F_i\mid Q) = \frac{w_i \cdot P_i(Q)}{\sum_j w_j \cdot P_j(Q)} $$
Here $w_i$ denotes the relevance mixing weight of expert $i$, and $P_i(Q)$ is the expert’s predictive distribution for query $Q$. The formula quantifies each expert’s contribution and dynamically adjusts weights across different queries, yielding a query‑specific relevance‑mixed FCM.\
We demonstrate the approach with synthetic FCMs generated from a real 13‑node expert FCM that models causes and feedbacks of ischemic‑stroke risk. Comparing equilibrium predictions before and after mixing shows that relevance weights markedly improve answer accuracy and robustness.\
Review: The paper ingeniously transfers the legal notion of relevance into the fuzzy cognitive map mixing framework. The Bayesian‑style de‑mixing provides an interpretable weight allocation for multi‑expert systems, and the stroke‑risk case study confirms its practical value.