Reinforced covariance weighted mean of vectors optimizer: insight, diversity, deep analysis and feature selection

The WeIghted meaN oF vectOrs (INFO) algorithm is widely used as an efficient optimization tool due to its simple structure and superior performance. However, achieving a balance in solving complex high-dimensional problems is difficult and quickly falls into premature convergence or local optimality. To better balance the conflict between exploration and exploitation capabilities, a Q-learning Covariance weIghted meaN oF vectOrs Algorithm (QCINFOCMA) based on reinforcement learning is designed in this study to solve the global optimization problem. QCINFOCMA incorporates a covariance matrix adaptation evolution strategy and Cauchy mutation as a new exploration scheme. The Q-learning strategy in reinforcement learning is also integrated into the original INFO to achieve adaptive switching between the original local search and the new exploration scheme. This allows search agents to use rewards and penalties to select exploration methods without following established models or strategies. In this study, a comprehensive analysis is conducted, pitting QCINFOCMA against 10 heuristics and 9 state-of-the-art algorithms, utilizing the IEEE CEC 2017 test functions. The experimental results show that QCINFOCMA outperforms other advanced algorithms in terms of convergence speed and convergence accuracy. Subsequently, QCINFOCMA was subjected to a discretization process, effectively transforming it into a binary tool through the application of a specific transformation function. This binary tool was then employed to address the real-world challenge of feature selection across a cohort of 36 datasets obtained from the UCI machine learning library. Empirical findings demonstrate that QCINFOCMA attains superior classification accuracy and requires fewer features in comparison to alternative optimization algorithms. The proposed QCINFOCMA can be a novel optimization tool for implementing global optimization and wrapper-based feature selection tasks.