A Hyperheuristic and Reinforcement Learning Guided Meta-heuristic Algorithm Recommendation

Automatic selection of the most appropriate algorithms for complex optimization problems has emerged as a cutting-edge trend in artificial intelligence. This approach circumvents the interpretability challenges posed through trial and error. A hyperheuristic and reinforcement learning-guided meta-heuristic algorithm recommendation (HHRL-MAR) is proposed to facilitate the adaptive selection of a diverse array of meta-heuristic algorithms tailored to the unique characteristics of various problems in this paper. To this end, four meta-heuristics with distinct advantages are integrated to form the action space within the reinforcement learning, serving as the low-level heuristic for hyperheuristic. The incorporated reward mechanism based on the real-time state of the population enhances both the flexibility and accuracy of the algorithm. Three selection strategies in light of simulated annealing and epsilon - greedy are designed to avoid premature convergence associated with a singular selection approach. The experimental results show the efficacy of HHRL-MAR for large-scale complex continuous optimization in terms of accuracy, stability, and convergence speed.