Aleatoric vs Epistemic Uncertainty

Standard ML distinction (Hüllermeier & Waegeman 2021) emphasised by Elyasaf et al. 2025 as central to safe self-modification in ABPSs:

Aleatoric — inherent randomness (irreducible noise in durations, outcomes, resources). Modeled with probability distributions.
Epistemic — lack-of-knowledge uncertainty (reducible by gathering more data or targeted sensing). Addressed with active learning, Bayesian inference.

Distinguishing the two determines the mitigation strategy: more data (epistemic) vs probabilistic hedging (aleatoric). Uncertainty quantification also gates learning-to-defer: when epistemic uncertainty is high an ABPS should escalate to a human decision-maker.

Communicating uncertainty to humans requires pairing quantitative measures with XAI-style explanations (Miller 2019: "probabilities probably don't matter"), e.g. via natural-language generation or Shapley-based explanations.

Relation to the philosophical probability debate

The aleatoric/epistemic split maps onto the chance/credence distinction in philosophy of probability (sources/2023-anjum-rocca-phi403-lecture-15-credence): aleatoric uncertainty is ontological probability (worldly chance or propensity); epistemic uncertainty is credence (degree of belief under limited knowledge). See concepts/probabilistic-causation for the philosophical context.

concepts/self-modification · concepts/explainability-apm · concepts/predictive-process-monitoring · concepts/probabilistic-causation