Automated Reasoning in the Age of the Internet

Abstract

The internet hosts a vast store of information that we cannot and should not ignore. It’s not enough just to retrieve facts. To make full use of the internet we must also infer new information from old. This is an exciting new opportunity for automated reasoning, but it also presents new kinds of research challenge.

• There are a huge number of potential axioms from which to infer new theorems. Methods of choosing appropriate axioms are needed.

• Information is stored on the Internet in diverse forms, e.g., graph and relational databases, JSON (JavaScript Object Notation), CSV (Comma-Separated Values) files, and many others. Some contain errors and others are incomplete: lacking vital contextual details such as time and units of measurements.

• Information retrieved from the Internet must be automatically curated into a common format before we can apply inference to it. Such a representation must be flexible enough to represent a wide diversity of knowledge formats, as well as supporting the diverse kinds of inference we propose.

• We can employ forms of inference that are novel in automated reasoning, such as using regression to form new functions from sets of number pairs, and then extrapolation to predict new pairs.

• Information is of mixed quality and accuracy, so introduces uncertainty into the theorems inferred. Some inference operations, such as regression, also introduce uncertainty. Uncertainty estimates need to be inherited during inference and reported to users in an intelligible form.

We will report on the FRANK (Formally know as RIF: Rich Inference Framework. We changed the name as the RIF acronym is already in use, standing for Requirements Interchange Format.) system that explores this new research direction.

This work has been funded by a University of Edinburgh studentship for the second author and Huawei grant HIRP O20170511.