Explainable Algorithm Selection and Configuration through Instance Space

4 minute read

Picure of futuristic data

This is one of two research projects developing new tools for automated algorithm selection and configuration. The University of Melbourne is the home institution for this project. To view the Manchester-based partner project, click here.

This project will develop new analysis methods to help explain the performance (or lack thereof) of the algorithms that are used daily to plan our deliveries, organise our manufacturing plants, schedule our bus routes, optimise our supply chains, etc.

By intelligently exploring the space of possible instances of a problem and analysing the behaviour of the algorithms for such instances, the techniques developed in this project will be able to explain to decision-makers under which conditions we can expect those algorithms to provide trustworthy solutions and when we may expect that the solutions provided will be infeasible or suboptimal.

Outcomes of the project include new tools for automated algorithm selection and configuration, evaluated on a series of real-world industrial optimisation problems.

Project goals

The goals of this project are to:

  1. Improve the explainability of optimisation algorithms.
  2. Develop tools to support the automatic configuration of the most suitable algorithms to solve many real-world problems arising in logistics, manufacturing, healthcare, and other fields.

The main research questions are:

  1. How can we improve the explainability of mathematical optimisation algorithms to better understand which problem features influence the choice of algorithmic parameters?
  2. How can we detect and correct biases in the design and configuration of optimisation algorithms arising from the particular problem instances that have historically been used for designing and benchmarking such algorithms?
  3. How can we generate new problem instances that add value to the analysis of explainability of such algorithms, and ensure the analysis is unbiased?
  4. How can we design automated tools to reveal, for a given problem instance, which algorithm is likely to be best and how its parameters should be configured to enable it to achieve optimal performance?

Supervision team

The University of Melbourne: Professor Kate Smith-Miles, Doctor Mario Andrés Muñoz-Acosta

*Click on the researcher's name above to learn more about their publication and grant successes.

The University of Manchester: Associate Professor Manuel López-Ibáñez, Professor Julia Handl

Who we are looking for

We are seeking a PhD candidate with the following skills:

  • Demonstrated experience in the field of computer science or mathematics and statistics.
  • Demonstrated experience with mathematical optimisation/machine learning.
  • Demonstrated ability to work independently and as part of a team.
  • Demonstrated ability to write research reports or other publications to a publishable standard (even if not published to date).
  • Excellent written and oral communication skills.
  • Demonstrated time management, project management and organisational skills, and ability to work to priorities.
  • Demonstrated problem-solving abilities.

Further details

The PhD candidate will benefit from the combined expertise of the project supervisors, and the embedding into two research environments. The team led by Melbourne Laureate Professor Kate Smith-Mileswill contribute by extending unique tools for Instance Space Analysis to analyse and visualise the relation between instance features and configuration criteria as well as how a stream of instances evolves over time. Associate Professor Manuel López-Ibáñez (PI) is a worldwide expert on automated configuration of algorithms and Professor Julia Handl is an expert in multi-objective optimisation and machine learning, including for streaming data.

This PhD project will be based at the University of Melbourne with a minimum 12-month stay at the University of Manchester.

The candidate will be enrolled in the PhD program at the School of Mathematics and Statistics at the University of Melbourne and in the PhD program at the Alliance Manchester Business School at the University of Manchester.

This project brings together the world-leading expertise found at University of Melbourne and University of Manchester to tackle these breakthrough ideas. The team at Melbourne, led by Melbourne Laureate Prof Kate Smith-Miles, have developed unique tools to visualise the “instance space” of optimisation problems, and understand for which types of instances an algorithm is likely to perform well or poorly.

Their methodology known as Instance Space Analysis is available in an online tool called the Melbourne Algorithm Test Instance Library with Data Analytics (MATILDA, https://matilda.unimelb.edu.au) developed with support from an Australian Research Council Laureate Fellowship (2014-2020). While MATILDA provides visual insights into the geometry of the problem and the impact on various algorithms, it does not currently provide explanations about the interplay with the algorithm configuration space.

The candidate will be co-supervised at Melbourne by Dr Mario Andrés Muñoz-Acosta, an expert on optimisation and machine learning, who has worked extensively on the development of MATILDA as the lead postdoctoral researcher on the team.


To apply for this joint PhD opportunity, and to view the entry requirements, visit How to apply.

Related items

happy graduate students

How to apply

Apply for a joint PhD with the University of Melbourne and the University of Manchester.

Flexible Material_Getty

Current projects

Discover what researchers from the Manchester-Melbourne Research Training Group are working on right now.

A scientist sitting in a dark room looking at an illuminated laser disk

Graduate researcher experience

Hear the stories of current and past graduate researchers. Find out about their experiences at the University and where their degrees have taken them.