Transfer Learning Based Dynamic Optimization Algorithm

The optimization problem is ubiquitous in the real world.There are many mature algorithms for general single-objective and multi-objective optimization problems.Howev-er,there is still a lot of room for development on such complex issues as dynamic multi-objective optimization and multi-task optimization problems.For dynamic multi-objective optimization problems,MT-DMOEA algorithm is pro-posed in this paper.The innovation of MT-DMOEA has two aspects.The first is the combination of the manifold-based transfer learning method with the manifold-based multi-objective optimization algorithm.The transfer learning method is used to predict the initial population at a new moment.The optimization algorithm is used to solve the multi-objective optimization problem.The second is the design of a new memory mechanism,which record-s the previous generations of optimal solutions in the external storage space and then used to predict the initial population at the new moment.MT-DMOEA merges the initial popula-tions predicted by the two methods,and then uses a multi-objective optimization method to solve the problem at the new moment.Experimental results show that the solution obtained by MT-DMOEA has better convergence and diversity than some popular algorithms.For multi-task optimization problems,MTOPS and MTOMOPS are proposed in this paper to solve single-objective multi-task optimization problems and multi-objective multi-task optimization problems.Both MTOPS and MTOMOPS are based on particle swarm op-timization.Particle self-replication and multi-directional evolution mechanisms are added,enabling the particles to evolve in multiple directions simultaneously and performing differ-ent tasks.These methods achieve information interaction meanwhile do not interfere with the optimal solution information already obtained in the original direction of the particle.Experiments show that MTOPS algorithm has higher convergence speed and accuracy than existing algorithms,and MTOMOPS algorithm also has higher convergence speed,and the solution obtained has higher diversity and convergence.Finally,this paper also studies a class of more complex optimization problems:dy-namic multi-objective multi-task optimization problems,then proposes d-MTOMOPS algo-rithm.d-MTOMOPS combines the results of the first two parts of the paper:using trans-fer learning methods to predict the initial population,and then use MTOMOPS to solve problems in the new environment.The experimental results show that the d-MTOMOPS algorithm has a high convergence speed and accuracy.The significance of the research in this paper is to propose corresponding optimization algorithms for the three optimization problems.The solution obtained by the proposed algorithm is better than the existing algorithm in terms of convergence speed and accuracy,and has practical application significance to the optimization problems.