Algorithm Selection for Large-Scale Multi-objective Optimization
The present study applies Algorithm Selection to automatically specify the suitable algorithms for Large-Scale Multi-objective Optimization. Algorithm Selection has known to benefit from the strengths on multiple algorithm rather than relying one. This trait offers performance gain with limited or no contribution on the algorithm and instance side. As the target application domain, Multi-objective Optimization is a realistic way of approaching any optimization tasks. Most real-world problems are concerned with more than one objective/quality metric. This paper introduces a case study on an Algorithm Selection dataset composed of 4 Multi-objective Optimization algorithms on 63 Large-Scale Multi-objective Optimization problem benchmarks. The benchmarks involve the instances of 2 and 3 objectives with the number of variables changing between 46 and 1006, Hypervolume is the performance indicator used to quantify the solutions derived by each algorithm on every single problem instance. Since Algorithm Selection needs a suite of instance features, 4 simple features are introduced. With this setting, an existing Algorithm Selection system, i.e. ALORS, is accommodated to map these features to the candidate algorithms’ performance denoted in ranks. The empirical analysis showed that this basic setting with AS is able to offer better performance than those standalone algorithms. Further analysis realized on the algorithms and instances report similarities/differences between algorithms and instances while reasoning the instances’ hardness to be solved.
