Reinforcement learning-based comprehensive learning grey wolf optimizer for feature selection

Turning heavy and high-dimensional raw data into knowledge for decision makers is a complex process. Feature selection (FS) can do this task well by removing irrelevant and/or redundant parts from the raw dataset, aiming at reducing dimensionality and improving accuracy. In this work, a reinforcement learning-based comprehensive learning grey wolf optimizer (RLCGWO) is designed to solve the FS problem, which is modeled as a combinatorial optimization problem. First, a comprehensive learning operator is proposed, containing the static and dynamic learning strategies. These two strategies provide GWO with the exploration capability in different ways. Second, a novel RL-based policy regulation technique is developed, which is based on the Q-learning framework. Individuals are considered as agents that obtain the state of the environment based on the state encoding technique. Meanwhile, agents select the most appropriate actions from the well-designed action set based on the information provided by the Q-table. Furthermore, agents update the stand-alone Q-table with rewards to provide themselves with more timely and accurate feedback. Third, a chaotic-based learning strategy is devised for leaders to improve the quality of the optimal solution. The comparison results of the proposed RLCGWO with six successful GWO variants and three typical algorithms on the benchmarks initially demonstrate its advantages in convergence speed and accuracy. The proposed RLCGWO is finally applied to the challenging FS problem. The comparison results with six popular algorithms on 15 UCI datasets and 3 real world high-dimensional datasets underscore its high adaptability and versatility. Taken together, the proposed RLCGWO is a promising technique for FS.