Representing and Combining Partially Specified CPTs
This work addresses a specific issue in probabilistic graphical models for researchers in AI and machine learning, but it appears incremental as it extends previous work on network fragments.
The paper tackles the problem of representing and combining partially specified conditional probability tables (CPTs) by introducing an asymmetry network and an object-oriented representation, showing that this representation is parsimonious and defining an algebra for factoring CPTs and combining networks.
This paper extends previous work with network fragments and situation-specific network construction. We formally define the asymmetry network, an alternative representation for a conditional probability table. We also present an object-oriented representation for partially specified asymmetry networks. We show that the representation is parsimonious. We define an algebra for the elements of the representation that allows us to 'factor' any CPT and to soundly combine the partially specified asymmetry networks.