基于自适应惩罚和多目标粒子群优化的AdaBoost方法OA

To solve the issue of noise samples causing bias in AdaBoost training and the difficulty of achieving stable pruning results with single-objective particle swarm optimization for pruning ensemble models,the AdaBoost method was proposed based on the adaptive penalty and multi-objective particle swarm optimization.The adaptive penalty strategy was proposed to apply weighted penalties to suspicious noise for reducing impact on the subsequent training process.The two-stage multi-strategy multi-objective particle swarm optimization algorithm(TSMSMOPSO)was introduced for ensemble pruning.During the search phase,the particles were accelerated toward non-dominated particles to avoid searching worthless space.The global optimum was selected by considering the trade-off between diversity and convergence.To prevent getting stuck in local optima,the reference particles were randomly mutated to generate elite particles for enhancing population diversity in the exploitation phase.To validate the performance of the proposed algorithm,seven comparison algorithms were evaluated on 16 datasets.The adaptive penalty strategy and TSMSMOPSO were verified through the ablation tests and the pruning comparison experiment.The results show that the proposed algorithm achieves the highest accuracy on 12 datasets and the best F,score on 13 datasets,where the differences from the suboptimal values are 0.19％-16.67％and 0.62％-16.67％,respectively.Compared to single-objective particle swarm optimization,the pruned ensemble model of TSMSMOPSO is lighter in ensemble size and exhibits more stable pruning effects.