Multi-criteria Automatic Algorithm Configuration under Streaming Problem Instances

This research project aims to:

• Develop automated configuration (AC) methods capable of generating a portfolio of configurations according to multiple criteria when facing a stream of problem instances
• Develop AC methods based on an instance space analysis that permits parameter tuning based on knowledge of where an instance lies in the instance space, how its characteristics necessitate re-configuration of the algorithm, and why
• Augment test suites to include more strategically chosen test instances to improve explainability

The details

Many critical problems in logistics, manufacturing, healthcare and other fields are solved by optimisation and machine learning algorithms. Thanks to advances in automatic configuration tools, we are now able to automatically tune the parameters of these algorithms for new problems with minimal human effort.

Unfortunately, these tools are designed to tune algorithms according to a single criterion and assume that the characteristics of a problem do not change over time. In the real world, however, the users of such algorithms often face conflicting criteria, such as the time required to solve the problem versus the expected quality of the solution returned by the algorithm.

Moreover, it is often the case that similar problems must be solved regularly (for example daily) in the case of a parcel delivery service, a manufacturing plant processing orders in daily batches or the daily planning of operating theatres in hospitals.

In those cases, the characteristics of the daily instances of the problem may evolve over time due to economical, societal and technological changes.

This project aims to extend the capabilities of automatic configuration tools to handle multiple conflicting criteria and adapt to such changes in the problem characteristics.

For this purpose, the teams at Manchester and Melbourne will join their expertise in the automatic configuration of algorithms and instance space analysis. The result of this project will be more powerful tools for tuning and deploying the critical algorithms that our modern world relies on so that they can better adapt to changes in the problems being solved and let users decide the most appropriate trade-off among conflicting criteria.

The graduate researcher on this project is: Javier Mora Jimenez

Supervision team

• The University of Manchester: Associate Professor Manuel López-Ibáñez, Professor Julia Handl
• The University of Melbourne: Professor Kate Smith-Miles, Doctor Mario Andrés Muñoz-Acosta