ESARSA-MRFO-FS: Optimizing Manta-ray Foraging Optimizer using Expected-SARSA reinforcement learning for features selection

Disease prediction with the help of computers has achieved significant progress in this area; however, it still requires a more accurate identification of each data feature. In the past few years, ML-based medical diagnoses have become increasingly dependent on the datasets rather than the algorithms. Consequently, the more features in a dataset, the more difficult the model training process will be. A feature selection (FS) strategy can be used to eliminate unwanted features before the training in order to improve the model performance. Due to the great increase in the number of features, FS causes a complex combinatorial problem, which is NP-hard. The primary goal of this research is to create a novel framework for improving the Manta-ray Foraging Optimizer (MRFO) algorithm’s performance. By the method of exploration–exploitation toggle adjustments, this process becomes the most efficient one by using the Expected State–action-reward-state–action (ESARSA) reinforcement learning strategy. The study employed 12 datasets as well as two classifiers (K-Nearest-Neighbor and Decision-Tree), both with and without hyperparameter adjustment. The study compared the performances of three feature selection methods: no feature selection, the original MRFO algorithm, and the proposed ESARSAMRFO-FS algorithm. The results revealed that ESARSA-MRFO-FS is the most effective of the three algorithms with classifiers and optimization scenarios. Besides, experiments comparing MRFO and ESARSA-MRFO performance on 44 benchmarks are presented in the article. In five functions, ESARSA-MRFO outperforms MRFO. These results confirm that the proposed method is efficient for feature selection in medical diagnosis because of its accuracy and low processing costs.