基于强化学习的多目标特征选择算法研究

RESEARCH ON MULTI-OBJECTIVE FEATURE SELECTION ALGORITHNM BASED ON RENFORCEMENT LEARNING

摘要:

特征选择作为降维方法，能够减少高维数据计算复杂度。然而传统方法多集中于单目标优化，难以同时平衡分类准确率和特征数量，特别是在高维数据中存在收敛性不足的问题。针对该问题，提出了一种基于强化学习和双种群策略的多目标特征选择算法（QSNSGA-Ⅱ），该算法通过强化学习动态调整关键参数，在保持种群多样性的同时加快收敛速度。采用双种群协同进化策略，增强了全局搜索能力，提高了算法性能。通过在多个UCI数据集上进行实验，尤其在高维数据集Musk1和Sonar数据集上准确率达到了91.1585%和84.636%，展示了该算法优异的特征选择效果及较好的收敛性和多样性。

Abstract:

Feature selection as a dimension reduction method can reduce the computational complexity of high-dimensional data. However, traditional methods mostly focus on single-objective optimization, which is difficult to balance the classification accuracy and the number of features at the same time, especially in high-dimensional data with the problem of insufficient convergence. To address this problem, a multi-objective feature selection algorithm (QSNSGA-II) based on reinforcement learning and dual-population strategy is proposed, which dynamically adjusts the key parameters through reinforcement learning to accelerate the convergence speed while maintaining the diversity of the population. A two-population co-evolutionary strategy is used to enhance the global search capability and improve the performance of the algorithm. Through experiments on several UCI datasets, especially on the high-dimensional datasets Musk1 and Sonar datasets, the accuracy reached 91.1585% and 84.636%, which demonstrated the excellent feature selection effect of the algorithm and better convergence and diversity.