Learning Hybrid Constraints and Using them to Reason Effectively

Hybrid constraints allow various reasoning tasks to be performed: probabilistic inference, predictions and optimisation. This proposal studies how these constraints can be learned from data and how those aforementioned reasoning tasks can be solved efficiently. Prior research has enabled 1) automatically learning constraints in spreadsheets with the ultimate goal of making predictive machine learning available to non-computer scientists, 2) learning constraints that are used in hybrid probabilistic inference and mathematical optimisation, and 3) working on developing solvers that can reason with these constraints efficiently. This proposal aims to cast predictive machine learning as hybrid probabilistic inference, broaden the expressivity of constraints that can be learned, improve existing learning and solving algorithms, and develop approximate solving algorithms.

Keywords:constraint learning, probabilistic inference, constrained reasoning, constrained prediction

Disciplines:Information technologies