A Two-Level Approach to Maximum Entropy Model Computation for Relational Probabilistic Logic Based on Weighted Conditional Impacts

The principle of maximum entropy allows to define the semantics of a knowledge base consisting of a set of probabilistic relational conditionals by a unique model having maximum entropy. Using the concept of a conditional structure of a world, we define the notion of weighted conditional impacts and present a two-level approach for maximum entropy model computation based on them. Once the weighted conditional impact of a knowledge base has been determined, a generalized iterative scaling algorithm is used that fully abstracts from concrete worlds. The weighted conditional impact may be reused when only the quantitative aspects of the knowledge base are changed. As a further extension of previous work, also deterministic conditionals may be present in the knowledge base, and a special treatment of such conditionals reduces the problem size.