Using deep Q-network for selection hyper-heuristics
Augusto Dantas, Alexander Fiabane do Rego, Aurora Pozo · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2021
Hyper-Heuristics is an active research field that aims to automatically select (or generate) the best low-level heuristic in each step of the search process. This work investigates a Hyper-Heuristic with a Deep Q-Network (DQN) selection strategy and compares it with two state-of-the-art approaches, namely the Dynamic MAB and the Fitness-Rate-Rank MAB. The experiments conducted on two domains from the HyFlex framework showed that the DQN approach outperformed the others on the Vehicle Routing Problem and was competitive on the Traveling Salesman Problem. This indicates that the DQN is a robust selection strategy that is less sensitive to the domain than the MAB based approaches.