A Reinforcement Learning Hyper-heuristic for the optimisation of Flight Connections

Many combinatorial computational problems have been effectively solved by means of hyper-heuristics. In this study, we focus on a problem proposed by Kiwi.com and solve this problem by implementing a Reinforcement Learning (RL) hyperheuristic algorithm. Kiwi.com proposed a real-world NP-hard minimisation problem associated with air travelling services. The problem shares some characteristics with several TSP variants, such as time-dependence and time-windows that make the problem more complex in comparison to the classical TSP. In this work, we evaluate our proposed RL method on kiwi.com problem and compare its results statistically with common random-based hyper-heuristic approaches. The empirical results show that RL method achieves the best performance between the tested selection hyper-heuristics. Another significant achievement of RL is that better solutions were found compared to the best known solutions in several problem instances.