|zpravodaj ČSKI - únor 2005 [ pdf ]|
datum: 8.2.2005 v 14:00
název: On learning of Bayesian network models by maximization of a quality criterion
přednášející: Milan Studený (UTIA)
místo konání: UTIA, místnost č.474
souhrn: The talk will concern with the question of learning graphical models of conditional independence structure, more specifically the question of learning Bayesian network models.
The first part of the talk will be a methodological overview of methods for structural learning of Bayesian networks based on maximization of a quality criterion. In particular, basic mathematical requirements on the respective quality criterion will be specified. Then, the idea of so-called local search method, which is commonly used method for maximization of a criterion of this kind, will be recalled.
The second part of the lecture will be devoted to a special arithmetic approach to this problem. The basic idea is to represent the respective Bayesian network model by a special arithmetic representative, namely a certain integral vector, called a standard inset. Advantages of this approach will be pinpointed and some open problems will also be formulated. Later results mentioned in the talk were achieved in cooperation with Jiří Vomlel, a colleague from the Department of Decision-Making Theory.