30 June 2025 in Proceedings of the Institution of Mechanical Engineers, Part P: Journal of Sports Engineering and Technology
Recent trends in Major League Baseball (MLB), including the “fly-ball revolution” and restrictions on defensive shifts, have increased the defensive burden on outfielders. This study aimed to optimize outfield defensive positioning using machine learning techniques, specifically k -means clustering and logistic regression. Analyzing 51,290 batted ball landing locations from the 2023 MLB season, we examined how the distance between an outfielder’s position and a batted ball’s landing spot affects the hit probability. Outfielder positions were reconstructed using trigonometric conversions of depth and angle, and three distance metrics – Euclidean, x -coordinate, and y -coordinate – were measured. The results indicate that shorter distances between an outfielder’s position and a batted ball’s landing spot significantly increase defensive success rates. Notably, k -means clustering produced optimal defensive positions closer to the ball in Euclidean and y -coordinate distances than actual outfielder positions. These findings provide practical insights that can assist coaches and analysts in optimizing defensive alignments, improving player positioning strategies, and enabling real-time defensive adjustments based on pitcher-batter matchups to enhance overall team defensive efficiency.
25 November 2025 in Proceedings of the Institution of Mechanical Engineers, Part P: Journal of Sports Engineering and Technology
This study investigated the impact of an automated ball-strike system (ABS) on baseball game dynamics and pitcher performance using a deep learning-based counterfactual approach. We developed deep neural network (DNN) models to predict game duration, walks and hits per inning pitched (WHIP), walks, and hits allowed using game-level data from the 2021–2023 Korea Baseball Organization seasons. Through a comparative analysis of three widely adopted machine learning algorithms, this study validated the suitability of DNN models for counterfactual evaluations of policy interventions in professional baseball. The model evaluation demonstrated high predictive accuracy, with mean absolute error values of 9.62 min for game duration, 0.205 for WHIP, 0.866 for walks, and 1.001 for hits allowed. Counterfactual predictions for the 2024 season, compared with observed outcomes, revealed statistically significant differences associated with ABS implementation. Game duration decreased, reflecting improved efficiency, whereas WHIP, walks, and hits allowed increased, suggesting a more challenging environment for pitchers. These findings highlight the dual impact of ABS: a shortened game duration and altered pitcher-batter dynamics. This study introduces a deep learning-based approach for evaluating policy changes in professional sports, demonstrating its utility in quantifying the effects of ABS technology on game outcomes. The results emphasize the need for further research on the long-term implications of ABS and its potential influence on strategic decision-making in baseball.
14 November 2025 in Proceedings of the Institution of Mechanical Engineers, Part P: Journal of Sports Engineering and Technology
Ball-strike judgments in Major League Baseball (MLB) remain susceptible to human error, potentially affecting game outcomes and competitive fairness. This study develops an AI umpire using deep learning to assess the consistency of human umpire decisions. A deep neural network trained on 362,734 pitches from the 2019 MLB season achieved an accuracy of 93%. The AI umpire’s high-confidence predictions (probability >0.9 or <0.1) were compared against actual umpire calls to identify systematic discrepancies. Results indicate that these discrepancies were not randomly distributed but concentrated in specific game contexts—particularly early innings and first-pitch situations—suggesting situational influences on judgment. Moreover, some umpires exhibited spatial biases by consistently expanding or contracting the strike zone relative to the AI model’s zone. These findings suggest that umpire decisions are shaped by game dynamics rather than being solely rule-bound. This study proposes a novel AI-based evaluation framework for officiating consistency, illustrating the potential of AI to support more objective and reliable assessments than human umpires.