A Double-Layer Reinforcement Learning Feature Optimization Framework for Evolutionary Computation Based Feature Selection Algorithms

Recently, Evolution Computing (EC) has gained widespread use in Feature Selection due to its powerful search capabilities. However, many EC algorithms fail to fully utilize historical combination information between features. Moreover, when faced with ultra-high dimensional data, they often lack the necessary decision-making ability to select a suitable optimization direction. In this paper, we propose a double layer-reinforcement learning framework feature optimizes framework. The framework aids the EC algorithm by continuously obtaining and utilizing combined feedback from features during the iteration process. We leverage the adaptability and decision-making abilities of reinforcement learning to overcome the EC algorithm’s limitations. We conducted experiments on 8 datasets from UCI to evaluate the effectiveness of our framework. The experimental results demonstrated that the EC algorithms, optimized by our framework, achieves lower error rates and requires fewer features. Consequently, we posit that reinforcement learning can offer novel methods and ideas for the application of evolutionary computing in feature selection.