Deep Dive: combinatorial optimisation with machine-learning based input

How can we benefit from the integration of both machine learning and reasoning? Get some great insights from two experts.

Machine learning (e.g., neural networks) and constraint programming are two core techniques in AI. These subfields have been developed mostly independently from one another. But what if we would combine the two? That’s exactly what professor Tias Guns (Flanders AI Research Program) and his team are exploring in their latest research.

When it comes to documents for example, scanning often isn’t enough. Let’s say the information on a subsidy request has to be verified, this goes beyond perception. Here interpreting will come into play. Machine learning models used for scanning often make mistakes here. But combine the machine learning with knowledge of the requirements and constraints and you’ll get a much better result.

To demonstrate the benefits of the integration of both machine learning and reasoning they developed a Sudoku Assistant app (available on Google Play store). The app uses machine learning as well as constraint programming and it seems it solves more sudokus correctly than previous apps were able to. Guns expects a lot of real-world issues will profit from this hybrid approach: for example solutions for vehicle routing, scheduling and rostering will be more desirable and understandable.

Want to know more:

Join us this Thursday, March 16, for a deep dive into constraint programming. Professors Pierre Schaus and Tias Guns will be reviewing different combinatorial optimisation problems as well as discrete optimisation techniques for solving them.

This course is part of a VAIA and Trail joint seminar series and is free of charge.

Don’t forget to subscribe!