Search Results (1)

In the development of artificial intelligence (AI) technology, utilizing datasets for model instruction to achieve higher predictive and reasoning efficacy has become a common technical approach. However, primordial datasets often contain a significant number of redundant features (RF), which can compromise the prediction accuracy and generalization ability of models. To effectively reduce RF in datasets, this work advances a new version of the Pufferfish Optimization Algorithm (POA), termed AMFPOA. Firstly, by considering the knowledge disparities among different groups of members and incorporating the concept of adaptive learning, an adaptive exploration strategy is introduced to enhance the algorithm’s Global Exploration (GE) capability. Secondly, by dividing the entire swarm into multiple subswarms, a three-swarm search strategy is advanced. This allows for targeted optimization schemes for different subswarms, effectively achieving a good balance across various metrics for the algorithm. Lastly, leveraging the historical memory property of Fractional-Order theory and the member weighting of Bernstein polynomials, a Fractional-Order Bernstein exploitation strategy is advanced, which significantly augments the algorithm’s local exploitation (LE) capability. Subsequent experimental results on 23 real-world Feature Selection (FS) problems demonstrate that AMFPOA achieves an average success rate exceeding 87.5% in fitness function value (FFV), along with ideal efficacy rates of 86.5% in Classification Accuracy (CA) and 60.1% in feature subset size reduction. These results highlight its strong capability for RF elimination, establishing AMFPOA as a promising FS method. Full article