<p>This study proposes a metaheuristic named ABC-GA-QL for the travelling salesman problems (TSP) where the Q-learning approach is used to select suitable metaheuristic methods in different iterations at run time from two different type metaheuristic methods-artificial bee colony (ABC) and genetic algorithm (GA). The algorithm involves two phases. The first phase explores the search space using ABC and GA independently to search the prominent zones of the optimal paths. The performances of GA and ABC on a target problem are noted at each iteration using a reward table (RT) and a Q-table (QT). The second phase uses the method with better performance in each iteration to modify the population. The entries of RT and QT are updated accordingly. In this way, ABC-GA-QL enhances the quality of the search process. The basic perturbation rule of ABC is modified for the TSPs using swap operation and the exploitation of the method is updated using 3-Opt operation. GA is also modified using multi-point cyclic crossover and 3-Opt operation to deal with the TSPs. ABC-GA-QL is tested against a set of asymmetric and symmetric TSPs from TSPLIB with significantly large sizes. The method finds optimal paths for the instances having sizes up to 280 nodes in most of the runs of the algorithm. Its performance is compared with some recently published state-of-the-art algorithms for the TSPs using the Friedman test, and the superiority of ABC-GA-QL is well established. The best results obtained in 10 independent runs of the method for different test instances are listed for deep statistical comparison in future studies. Some approaches are also proposed to solve the TSPs with cost matrices involving fuzzy data. The algorithm is also used to implement a real-life smart home delivery system for the online business.</p>