Based on the conventional genetic algorithm, this work suggests an adaptive variation genetic algorithm (AGA), which increases population diversity and speeds up convergence by increasing variation probability. The entropy approach is used to first assess the caliber of physical education and to produce a priori assessment samples. It is combined with the AGA-BP model based on adaptive variation probability. The BP neural network is then utilized for assessment learning in the field of evaluating the quality of physical education. It is optimized by an adaptive variation-based genetic algorithm. Finally, student physical activity levels were assessed both before and after the physical education reform using a more thorough and scientifically based EM-AGA-BP teaching quality evaluation model. The findings revealed that, at 41.2% and 53.4%, respectively, the percentage of students’ static activity time after the reform was much lower than that before the reform. By using independent samples t-tests, each revealed significant differences (P 0.05).