Estymacja prawdopodobieństwa subiektywnego w modelowaniu ryzyka

W analizach ryzyka spotykamy się z brakiem danych obiektywnych do estymacji rozkładów prawdopodobieństwa parametrów modeli. W takiej sytuacji wspomagamy się rozkładami subiektywnymi, estymowanymi na podstawie sądów eksperckich. Przedstawiona została charakterystyka błędów popełnianych przez człowieka szacującego prawdopodobieństwo. Dokonano przeglądu metod stosowanych w estymacji prawdopodobieństwa subiektywnego w analizach ryzyka. Zwrócono uwagę na metody neuronowe.

---

We can notice the lack of objective data for the estimation of model parameter probability distributions in risk analyses. In such situation we can aid the modelling by subjective distributions, which are estimated on the basis of experts' judgments. The characteristic of errors made by man when he is estimating probability and the taxonomy of probabilistic judgments used in risk analysis have been presented. The review of the subjective estimation methods used in risk analyses has been done. The methods, which seem particularly suitable for the risk modelling, are described more closely. Those are the methods based on the following models: Mendel and Sheridan (Bayesian model), classical by Cooke (non-Bayesian model) and pair comparison. The simple and neural methods are described as well. Lack of skill in probabilistic judgments by experts was assumed in the methods called by the author as simple. It is required from experts that they give only the range of values in which by their opinions the estimated value is contained and that value is taken from their professional praxis (it is not a parameter of probability distribution). Attention was paid to neural methods, which allow us to avoid the complex modelling of experts' credibility evaluation and combining of experts' judgment. They are simple and have given good results in the test estimation of sea voyages duration. In closing remarks it was underlined that there exist a number of efficacious methods of subjective probability estimation. Those can be used to aid the risk modelling created by engineering objects.