Reinforcement learning assisted sparse population coevolutionary algorithm for multi-component spectral feature selection

Abstract

As an essential step in spectral quantitative analysis, spectral feature selection identifies the most relevant and significant features from high-dimensional spectral data. This process aims to improve the accuracy of concentration prediction models while also reducing model complexity. However, existing evolutionary algorithms fail to account for the potential cooperation in this problem, which may degrade performance. This paper proposes a sparse population coevolutionary algorithm based on deep reinforcement learning for multi-component spectral feature selection. It introduces auxiliary sparse populations for single-component spectral feature selection and utilizes the Deep Q-learning Network (DQN) to select a population as an evolutionary helper, thereby accelerating the exploration and exploitation of the main sparse population for multi-component spectral feature selection. DQN establishes a mapping from population states to the selection action of an auxiliary population used for coevolution. The best auxiliary evolutionary population is selected based on the current state of the main population at each generation, thus promoting convergence towards the Pareto-optimal fronts. In the experiments, the meat and flue gas datasets are used to evaluate the effectiveness of the proposed algorithm. Experimental results indicate that the proposed algorithm is superior for multi-component spectral feature selection over four state-of-the-art evolutionary algorithms.