A transfer learning-based particle swarm optimization algorithm for travelling salesman problem

Abstract To solve travelling salesman problems (TSPs), most existing evolutionary algorithms search for optimal solutions from zero initial information without taking advantage of the historical information of solving similar problems. This paper studies a transfer learning-based particle swarm optimization (PSO) algorithm, where the optimal information of historical problems is used to guide the swarm to find optimal paths quickly. To begin with, all cities in the new and historical TSP problems are clustered into multiple city subsets, respectively, and a city topology matching strategy based on geometric similarity is proposed to match each new city subset to a historical city subset. Then, on the basis of the above-matched results, a hierarchical generation strategy of the feasible path (HGT) is proposed to initialize the swarm to improve the performance of PSO. Moreover, a problem-specific update strategy, i.e. the particle update strategy with adaptive crossover and clustering-guided mutation, is introduced to enhance the search capability of the proposed algorithm. Finally, the proposed algorithm is applied to 20 typical TSP problems and compared with 12 state-of-the-art algorithms. Experimental results show that the transfer learning mechanism can accelerate the search efficiency of PSO and make the proposed algorithm achieve better optimal paths.