Multitask Genetic Programming Based on Shapley Additive Explanation and Seismic Building Vulnerability Assessment Case

In traditional multitask genetic programming (MTGP), expressions for each task are selected independently based on certain criteria from the population. However, valuable knowledge embedded in expressions may be lost by disregarding those with lower accuracy. Inspired by Shapley additive explanation (SHAP) analysis, a SHAP analysis-aided MTGP algorithm, termed SHMTGP, is proposed to integrate valuable knowledge dispersed within the population. Specifically, a SHAP-guided selection method is designed as the core composition of SHMTGP, aiming to extract valuable knowledge from the population of each task. Experimental results on three groups of test functions demonstrate the superiority of SHMTGP over state-of-the-art methods. Additionally, a case study on seismic building vulnerability assessment further evidences the superiority of SHMTGP.