Transfer Learning Based Particle Swarm Optimization Algorithms Research And Applications

Particle swarm optimization is a kind of evolutionary optimization technique that simulates the evolutionary behavior of nature,and derived from the foraging behavior of birds.With the rapid development of evolutionary optimization technology,particle swarm optimization plays an increasingly important role in scientific research and engineering practice.However,it is still an open problem about how to speed up the convergence speed of the population and how to improve the search efficiency of the algorithm in the real optimization problems.It is found that the most of existing particle swarm optimization algorithms search for the optimal solution from the zero initial information of the problem,which wastes the population computing resources to some extent.Transfer learning is a machine learning method for solving problems in related fields that uses existing historical knowledge.It is of great research value to extend transfer learning to the field of evolutionary optimization,which can not only reduces the computational cost of the algorithm,but also improves the search ability of the population significantly.In view of this,this thesis studies particle swarm optimization algorithms and applications based on transfer learning,which mainly includes the following two parts:(1)Based on an evolutionary optimization framework of transfer learning with similar historical information,a particle swarm optimization algorithm based on transfer learning of similar historical information is proposed.It can improve the population's search efficiency of transferring the knowledge that corresponds to the historical problem to the solving process of the new problem,where the historical problem,matching with the new problem,is found in the historical model library of the solved problems.Firstly,define a maximum mean difference indicator based on multidistribution estimation to evaluate the matching degree between the new problem and historical models.A population initialization strategy based on model matching similarity is given to speed up the search efficiency of the population.A representative individual retention strategy based on rapid clustering is provided for updating the model library.Secondly,the ABPSO algorithm is embedded into the proposed framework,and a BBPSO algorithm based on transfer learning of similar historical information is presented.Finally,the proposed algorithm is applied to multiple standard test functions to verify the effectiveness of the experimental results.(2)Applying the idea of transfer learning to the evolutionary solution of the traveling salesman problem,a fast particle swarm optimization algorithm for largescale TSP problem under the guidance of transfer learning is proposed.Firstly,neighboring cities are clustered using the K-mediods technology,and a city topological matching strategy based on geometric angular similarity is defined to find historical city subsets which are similar to the city subsets to be optimized in distribution.Secondly,a feasible path generation strategy merged intra-class city sequential transfer learning with inter-class sequential greedy learning is proposed to initialize the particle swarm.Thirdly,an integer type update strategy of the particle based on adaptive crossover and mutation is designed to adapt to the characteristics of TSP problem.Finally,the results on 16 typical TSP test problems show that the effectiveness of the proposed method.