An Artificial Intelligence-Based Expert System with Grid-Search Calibration for the Execution Analysis of the Tippelt Skill on Parallel Bars
Main Article Content
Abstract
Execution scoring in men's artistic gymnastics remains heavily dependent on subjective human judgement, despite the existence of explicit deduction rules in the Fédération Internationale de Gymnastique (FIG) Code of Points. This study proposes and validates a two-tier, rule-based expert system for automated deduction estimation on the Tippelt skill performed on parallel bars. The system's knowledge base combines (a) deduction thresholds derived directly from the FIG Code of Points, Article 14.3, for biomechanical indicators explicitly covered by the regulation, and (b) thresholds for two additional indicators lacking an explicit numerical standard, obtained through a grid-search calibration procedure. To avoid circularity, calibration was performed exclusively on a training subset (n = 56, 80%) of 70 recorded attempts by a single elite male gymnast, and validated on an independent holdout subset (n = 14, 20%) that played no role in threshold selection. Calibration improved overall agreement with the mean score of five certified judges, reducing mean absolute error from 0.476 to 0.309 points and increasing the Pearson correlation from 0.887 to 0.914. Performance on the independent holdout set (MAE = 0.272, r = 0.913) closely matched full-sample performance, indicating that the calibrated thresholds generalize rather than overfit the calibration data. A paired-samples t-test revealed a small but statistically significant systematic bias (mean difference = -0.185, t(69) = -4.564, p < .001, Cohen's d = 0.546), with the expert system tending to slightly overestimate execution scores relative to human judges. These findings suggest that a transparent, rule-grounded expert system can approximate expert judging with reasonable fidelity while remaining fully interpretable, and may serve as a decision-support tool rather than a replacement for human officiating.
