An Improved Particle Swarm Optimization Algorithm Based Computer Unified Device Architecture to Symbolic Regression

Symbolic regression, as a data-driven function modeling method based on evolutionary computation, is widely applied in many fields. For scenarios that require real-time modeling, the timeliness of the algorithm becomes particularly important. This paper proposes a parallel computing approach utilizing the computational power of GPUs to parallelize the particle swarm optimization algorithm for solving symbolic regression problems. By introducing and Scatter operator and mutation operator and a hierarchical self-adaptive learning strategy of particles, the local exploration and global search capabilities of particles are enhanced to improve the efficiency of the algorithm. Numerical experimental results demonstrate that the proposed method achieves an acceleration ratio of approximately 20 times compared to running on a CPU, and the acceleration effect becomes more evident with the increase of the data scale.