Systematic construction of algorithm portfolios for a Maintenance Scheduling Problem

We investigate how combinations of evolutionary algorithms (portfolios) can be constructed efficiently for solving optimization problem instances drawn from a distribution. We consider selection methods ranging in intricacy and based on different principles including a technique for efficient tuning of metaheuristics, a racing algorithm. The selectors are used here to optimize instances of the Preventive Maintenance Scheduling Problem (PMSP) in power generation. Experiments show different behaviors of selectors in term of computational time needed to choose constituent algorithms and the performance of the generated portfolios at optimizing previously unseen PMSP instances. The racing selector offers a good trade-off between the computational time and the performance.