Abstract:

In double-layer capacitated arc routing problems (CARP), both the high-level configuration problem and the low-level service problem are considered as optimization objects. Aiming to the double-layer CARP problems, a learnable genetic algorithm (LGA) is proposed. The basic architecture of LGA is constructed, and two different forms of knowledge are designed to provide an important foundation for the LGA. In LGA, two extended heuristic approaches are employed to generate initial populations, the performance knowledge of operators is used to select a proper operator for the selection, crossover and mutation, the component knowledge is employed to select a proper breakpoint location for each crossover and mutation, and the partial replacement procedure is used to maintain population diversity. To validate the performance of LGA, 24 benchmark problems are solved by LGA and some extended heuristic methods. The LGA architecture will act as a helpful reference to the improvement of optimization approaches.