"Black box" optimisation, speciality of the new RandOpt team

Today, the aim is to develop digital optimisation methods in order to solve difficult "black box" optimisation problems and which address current issues affecting industry, biology, physics, chemistry and medicine.

Randomized Optimisation or "black box" optimisation

The RandOpt team is working on the elaboration of adaptive stochastic non-derived optimisation algorithms, i.e. optimisation algorithms which, on the one hand, depend on uncertainties and, on the other hand, do not use the derivatives of the function to be optimised. The term "black box" is used as this function is seen by the algorithm as a "black box", in that it cannot exploit its contents. "It is the typical context of problems in industry where we do not have access to the source code, just an executable version of the code, that is to say of the function to be optimised" , Anne Auger, head of the RandOpt team, specifies.

Algorithms, a step ahead of theory

The team's work consists in developing better algorithms and new optimisation methods for non-convex optimisation in the following contexts:

• large-scale optimisation, i.e. when the problems to be optimised have a large number of variables;
• constrained optimisation, which require taking constraints that are essential for the computation of solutions into consideration;
• multi-objective optimisation, where the wish is to optimise not one - but several - contradictory criteria at the same time, such as, for example, minimising costs whilst also maximising the quality or service life of a product.

In this field, algorithmic developments and practice are ahead of theory. In parallel to algorithmic development, the team has set itself a priority of developing and improving the theoretical aspects of these methods. "If the CMA-ES method, Covariance Matrix Adaptation Evolution Strategy, developed by Nikolaus Hansen, is already considered to be state-of-the-art in randomized optimisation, the team's aim is to take it to benchmark level over the next 15 years by developing a very strong mathematical theory to this aim" Anne Auger explains.

Finally, the team is also working on the updating and improvement of the COCO platform (COmparing Continuous Optimisers) created by its members in 2008. This open source platform aims to automate optimisation algorithm performance comparative analysis tasks: from the collection of benchmarking data to its processing and visualisation.

Numerous applications in industry, biology, physics, chemistry or medicine

If the work carried out by RandOpt forms, first and foremost, part of a fundamental research perspective, the results are already striking numerous chords in industry, aerospace, biology, chemistry or medicine. 

For example, the team has collaborated on launcher sizing problems or design issues with photovoltaic panels.

Biologists, physicists and chemists can also be confronted with problems concerning the calibration of models or parameters. Faced with this type of situation, the methods developed by RandOpt make it possible to optimise the work of these scientists by adjusting the model's data using experimental data.