Revistes Catalanes amb Accés Obert (RACO)

Knwoledge revision in Markov networks

Jörg Gebhardt, Christian Bogerlt, Rudolf Kruse, Heinz Detmer

Resum


A lot of research in graphical models has been devoted to developing
correct and eficient evidence propagation methods, like join tree propagation
or bucket elimination. With these methods it is possible to condition the
represented probability distribution on given evidence, a reasoning process
that is sometimes also called focusing. In practice, however, there is the
additional need to revise the represented probability distribution in order
to reflect some knowledge changes by satisfying new frame conditions. Pure
evidence propagation methods, as implemented in the known commercial
tools for graphical models, are unsuited for this task. In this paper we develop
a consistent scheme for the important task of revising a Markov network so
that it satisfies given (conditional) marginal distributions for some of the
variables. This task is of high practical relevance as we demonstrate with
a complex application for item planning and capacity management in the
automotive industry at Volkswagen Group.

Text complet: HTML