Investigating Learning Methods in an Architecture for Metaheuristic Hybridization Applied to Multi-Level Decision Problems
Metaheuristics Hybridization, Matheuristics, Oppositional Learning, Orthogonal Learning, Traveling Car Renter with Passengers.
This work extends the capabilities of a multi-agent architecture for hybridizing metaheuristics to include learning and Mathematical Programming techniques. The learning techniques addressed are the Orthogonal Project and Opposition. The way of applying such techniques is innovative, contemplating the learning of agents regarding the choice of heuristics to be applied at different moments along the search. This approach is compared to the use of those learning techniques according to their traditional application mode verified in works in the literature. Another contribution is the inclusion of Mathematical Programming techniques, which produce matheuristic algorithms. The use of Mathematical Programming methods is also an innovative element since few architectures for hybridizing metaheuristics contain such a resource. This work proposes a form of hierarchical hybridization for Combinatorial Optimization problems that have multiple decision levels. The algorithmic proposals are tested in the Traveling Car Renter Problem with Passengers. This problem, which belongs to the NP-hard class, requires decision-making at three different levels: route, car type and meeting demand for rides. The results of experiments are reported for three classes of instances, in a total of ninety-nine test cases with sizes ranging from 4 to 80 cities, 2 to 5 vehicles and 10 to 240 people demanding transportation.