Landscape-Aware Selection of Metaheuristics for the Optimization of Radar Networks
Quentin Renau · HAL (Le Centre pour la Communication Scientifique Directe) · 2022
Radar networks are complex systems that need to be configured to maximize their coverage or the probability of detection of a target.The optimization of radar networks is a challenging task that is typically performed by experts with the help of simulators.Alternatively, black-box optimization algorithms can be used to solve these complex problems.Many heuristic algorithms were developed to solve black-box optimization problems and these algorithms exhibit complementarity of performance depending on the structure of the problem.Therefore, selecting the appropriate algorithm is a crucial task.The objective of this CIFRE PhD is to perform a landscape-aware algorithm selection of metaheuristics in order to optimize radar networks.The main contributions of this PhD thesis are twofold.In this thesis, we define six properties that landscape features should satisfy and we study to what degree landscape features satisfy these properties.One of the six properties is the invariance to the sampling strategy.We found that, surprisingly to what was recommended in the literature, the sampling strategy actually matters.We found important discrepancies in the feature values computed from different sampling strategies.Overall, we found that none of the features satisfy all defined properties.These features represent the core of a landscape-aware algorithm selection.We applied the landscape-aware algorithm selection of metaheuristics on the optimization of radar network use-cases.On this use-cases, algorithms have similar performances and the gain to perform an automated selection of algorithms is small.Nevertheless, the performance of the landscape-aware algorithm selection of metaheuristics is similar to the performance of the single best solver (SBS).