| 07 |
Evaluating uncertainty in the probability estimationVariance computation for probabilistic inference outcome 
|
|---|
In probabilistic inference, which predicts the likelihood of an event based on data and the relationships among events, a lack of data increases the uncertainty of the prediction, affecting decision-making depending on it. This study introduces a technique for quantitatively computing the "variance of inference outcome," an indicator of this uncertainty. It has been difficult to compute due to the enormous computational burden. To address this, we use the technique of counting the number of solutions that satisfy given logical formulas. Our study can be used to make more robust decisions that could otherwise lead to serious consequences. For example, in predicting availability of infrastructure equipment, even if the prediction satisfies safety standards, a large variance value indicates high uncertainty. In such cases, our approach enables operational decisions such as encouraging additional data collection to avoid excess failure rate that exceeds the prediction.
[1] K. Nakamura, M. Nishino, N. Yasuda, “Variance computation for weighted model counting with knowledge compilation approach”, in Proc. The 40th Annual AAAI Conference on Artificial Intelligence (AAAI ’26), 2026.
Kengo Nakamura, Linguistic Intelligence Research Group, Innovative Communication Laboratory