Andres R. Masegosa, Serafin Moral

---

A New Framework for Learning Generalized Credal Networks

Abstract

In this paper we give a formal specification of the problem of learning credal networks from observations. It is based on considering different equivalent sample sizes for the Dirichlet prior distributions about the probabilities of the conditional distributions. The novelty is that we specify what is the set of possible decisions and that this set may also include the selection of the equivalent sample size from a set of observations. Different Bayesian approaches can be considered as particular cases of this general framework. Approximate and exact algorithms based on A$^*$ search procedure are provided to compute the set of undominated decisions. Some preliminary experiments are reported.

Keywords

credal networks, learning, imprecise sample size Dirichlet model, search algorithms

---

Download area

The paper is available in the following formats:

• pdf file

E-mail addresses

Andres R. Masegosa		andrew@decsai.ugr.es
Serafin Moral		smc@decsai.ugr.es