Rugby Defense Strategy Prediction Using Deep Learning

Abstract

Abstract The integration of data analytics in sports has revolutionized performance evaluation and strategic planning, yet rugby union remains underrepresented in this domain, particularly concerning defensive strategies. This thesis addresses this gap by developing a deep learning framework to classify and evaluate three prevalent defensive formations in rugby union: blitz, drift, and umbrella. Utilizing manually annotated positional data from Rugby World Cup matches, the study employs Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNN), and hybrid CNN-LSTM architectures to capture the spatio-temporal dynamics of defensive plays. The research further incorporates success prediction as a secondary task, assessing the e!ectiveness of each defensive strategy. To enhance model generalization, rugby-specific data augmentation techniques, including coordinated jittering and mirroring, are applied.Experimental results show that the hybrid CNN-LSTM model achieved the highest performance, reaching an overall accuracy of 97.22%, with strong strategy recognition and success prediction capabilities across augmented datasets.The findings demonstrate the potential of deep learning models to automate the classification of defensive formations and predict their success, o!ering valuable insights for coaches and analysts. This work contributes to the advancement of rugby analytics by introducing a scalable, objective approach to defensive strategy evaluation.