Symbolic regression (SR) aims to discover mathematical expressions from data that describe the relationship between input variable and the output variable, and genetic programming (GP) has shown strong potential for this task since it can automatically search for optimal solutions without requiring a predefined structure of the expression. Semantic GP further enhances GP by incorporating semantic information into the evolutionary process to provide behavior aware search guidance. However, existing semantic GP methods still face several challenges, including the issue of program bloat; high-dimensional semantic space; and requiring representative selection among semantically equivalent individuals. Therefore, several important aspects like bloat control, genetic operators, and ensemble learning in semantic GP need further investigation. This study aims to explore new semantic GP approaches to enhance predictive performance, interpretability, and search effectiveness. By introducing innovative subtree complexity control, adaptive training-instance selection, semantic-based genetic operators, and ensemble learning mechanisms, the proposed methods are expected to outperform state-of-the-art learning algorithms on regression or symbolic regression problems.