Evolutionary algorithm for learning Bayesian structures from data

In this paper we report an evolutionary approach to learning Bayesian networks from data. We explain reasons, which advocate such a non-deterministic approach. We analyze weaknesses of previous works and come to conclusion that we should operate in the search space native for the problem i.e. in the space of directed acyclic graphs instead of standard space of binary strings. This requires adaptation of evolutionary methodology into very specific needs. We propose quite new data representation and implementation of generalized genetic operators and then we present an efficient algorithm capable of learning complex networks without additional assumptions. We discuss results obtained with this algorithm. The approach presented in this paper can be extended with the possibility to absorb some suggestions from experts or obtained by means of data preprocessing