Probabilistic Reasoning As Symbolic Evaluation (PRAiSE) | SRI International

# Probabilistic Reasoning As Symbolic Evaluation (PRAiSE)

SRI researchers are developing a new probabilistic reasoning system that explores an important direction in advancing computers’ abilities to understand data, manage results and infer useful insights. SRI’s Probabilistic Reasoning As Symbolic Evaluation (PRAiSE) is a probabilistic programming paradigm that enables a more usable way to manage uncertain information.

Probabilistic methods, used to analyze and manage large volumes of information, have contributed important progress for machine learning. But traditional methods rely on encoding knowledge as graphical models limited to very simple types of representations such as tables and weight matrices, which cannot cope with applications involving very complex models and large data sets. Moreover, these representations do not retain the high-level structure of the original knowledge, making it much harder to explain conclusions in understandable terms.

Typical probabilistic methods either ground the model—that is, instantiate its constructs exhaustively into a much larger low-level representation—or sample through parts of it. For example, when processing an election model involving a large population of voters, a typical graphical model system would separately instantiate and reason about millions of individual voters, losing sight of the general rules that apply to entire populations. This makes the system both less efficient and less capable of explaining its conclusions succinctly.

PRAiSE uses symbolic, lifted inference that allows the specification of models in a higher-level language similar to those of a programming language (hence, “probabilistic programming”). PRAiSE retains this high-level representation throughout its processing. In the same election model, this means applying high-level rules about entire populations at once by considering groups of homogenous voters, as opposed to reasoning about voters individually. This improves inference and learning and makes the explanation of conclusions more comprehensible to humans.

The current library and demo is capable of performing probabilistic inference on models defined with the following constructs:

• propositional variables
• equality on categorical types
• equality and inequalities over bounded integers (more specifically, difference arithmetic).
• linear real arithmetic

In the future, other constructs will be included:

• algebraic data types
• random functions (also described as relational variables, or uninterpreted functions - the type of inference common in the lifted first-order probabilistic inference literature).

PRAiSE has been developed mostly under DARPA’s Probabilistic Programming for Advanced Machine Learning (PPAML) program, which sought to develop probabilistic programming to create more economical, robust and powerful applications that need less data to produce more accurate results.

PRAiSE has been applied in a new DARPA program, “World Modelers,” which seeks to model geopolitical situations through probabilistic modeling and reasoning. The reasoning module will be applied to an expert model to combine data from multiple sources including satellite images, crop and poverty prediction, temperature and rainfall forecasts.

We gratefully acknowledge the support of the Defense Advanced Research Projects Agency (DARPA) Probabilistic Programming for Advanced Machine Learning Program under Air Force Research Laboratory (AFRL) prime contract no. FA8750-14-C-0005; and the DARPA World Modelers Program under Army Research Office (ARO) prime contract no. W911NF-18-C-0012.  Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the view of DARPA, AFRL, ARO, or the U.S. government.