Compiling AI Engineering Models for Probabilistic Inference

Abstract

In engineering domains, AI decision making is often confronted with problems that lie at the intersection of logic-based and probabilistic reasoning. A typical example is the plan assessment problem studied in this paper, which comprises the identification of possible faults and the computation of remaining success probabilities based on a system model. In addition, AI solutions to such problems need to be tailored towards the needs of engineers. This is being addressed by the recently developed high-level, expressive modeling formalism called probabilistic hierarchical constraint automata (PHCA).

This work introduces a translation from PHCA models to statistical relational models, which enables a wide array of probabilistic reasoning solutions to be leveraged, e.g., by grounding to problem-specific Bayesian networks. We illustrate this approach for the plan assessment problem, and compare it to an alternative logic-based approach that translates the PHCA models to lower-level logic models and computes solutions by enumerating most likely hypotheses. Experimental results on realistic problem instances demonstrate that the probabilistic reasoning approach is a promising alternative to the logic-based approach.