Large-scale optimisation problems are usually hard to solve optimally. Approximation algorithms such as metaheuristics, able to quickly find sub-optimal solutions, are often preferred. This thesis focuses on multi-objective local search (MOLS) algorithms, metaheuristics able to deal with the simultaneous optimisation of multiple criteria. As many algorithms, metaheuristics expose many parameters that significantly impact their performance. These parameters can be either predicted and set before the execution of the algorithm, or dynamically modified during the execution itself.
While in the last decade many advances have been made on the automatic design of algorithms, the great majority of them only deal with single-objective algorithms and the optimisation of a single performance indicator such as the algorithm running time or the final solution quality. In this thesis, we investigate the relations between automatic algorithm design and multi-objective optimisation, with an application on MOLS algorithms. We first review possible MOLS strategies ans parameters and present a general, highly configurable, MOLS framework. We also propose MO-ParamILS, an automatic configurator specifically designed to deal with multiple performance indicators. Then, we conduct several studies on the automatic offline design of MOLS algorithms on multiple combinatorial bi-objective problems. Finally, we discuss two online extensions of classical algorithm configuration: first the integration of parameter control mechanisms, to benefit from having multiple configuration predictions; then the use of configuration schedules, to sequentially use multiple configurations.