Adaptive Operator Selection at the Hyper-level

Whenever a new problem needs to be tackled, one needs to decide which of the many existing metaheuristics would be the most adequate one; but it is very difficult to know their performance a priori. And then, when a metaheuristic is chosen, there are still its parameters that need to be set by the user. This parameter setting is usually very problem-dependent, significantly affecting their performance. In this work we propose the use of an Adaptive Operator Selection (AOS) mechanism to automatically control, while solving the problem, (i) which metaheuristic to use for the generation of a new solution, (exemplified here by a Genetic Algorithm (GA) and a Differential Evolution (DE) scheme); and (ii) which corresponding operator should be used, (selecting among five operators available for the GA and four operators for DE). Two AOS schemes are considered: the Adaptive Pursuit and the Fitness Area Under Curve Multi-Armed Bandit. The resulting algorithm, named as Adaptive Hyper-Heuristic (HH), is evaluated on the BBOB noiseless testbed, showing superior performance when compared to (a) the same HH without adaptation, and also (b) the adaptive DE and GA.