Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MuLMINet: Multi-Layer Multi-Input Transformer Network with Weighted Loss

Minwoo Seong, Jeongseok Oh, SeungJun Kim

TL;DR

MuLMINet tackles forecasting future badminton rally actions from early-stroke history by introducing a Multi-Layer Multi-Input Transformer that extends ShuttleNet with dual encoders and a Position Aware Gated Fusion decoder, enabling joint prediction of shot type and landing area coordinates. The method employs a weighted loss $L_T$ to balance type and location tasks via $L_T = α (L_{ST} + L_{SL}) + (1-α)(L_B + L_A + L_H + L_{PL} + L_{OL})$ and uses a Loss Selection Module to search across hyperparameters, achieving strong performance on the ShuttleSet-derived dataset. Feature selection guided by Cramer's V identifies eight informative inputs, including aroundhead, backhand, landing height, and player/opponent area coordinates, with a 5-fold CV framework and 300 training epochs. The approach culminates in a runner-up finish at IJCAI CoachAI Badminton Challenge Track 2 and is supported by publicly available code, highlighting practical potential for AI-assisted coaching and strategic planning in turn-based sports, while suggesting future work on correlation-aware embedding strategies.

Abstract

The increasing use of artificial intelligence (AI) technology in turn-based sports, such as badminton, has sparked significant interest in evaluating strategies through the analysis of match video data. Predicting future shots based on past ones plays a vital role in coaching and strategic planning. In this study, we present a Multi-Layer Multi-Input Transformer Network (MuLMINet) that leverages professional badminton player match data to accurately predict future shot types and area coordinates. Our approach resulted in achieving the runner-up (2nd place) in the IJCAI CoachAI Badminton Challenge 2023, Track 2. To facilitate further research, we have made our code publicly accessible online, contributing to the broader research community's knowledge and advancements in the field of AI-assisted sports analysis.

MuLMINet: Multi-Layer Multi-Input Transformer Network with Weighted Loss

TL;DR

MuLMINet tackles forecasting future badminton rally actions from early-stroke history by introducing a Multi-Layer Multi-Input Transformer that extends ShuttleNet with dual encoders and a Position Aware Gated Fusion decoder, enabling joint prediction of shot type and landing area coordinates. The method employs a weighted loss to balance type and location tasks via and uses a Loss Selection Module to search across hyperparameters, achieving strong performance on the ShuttleSet-derived dataset. Feature selection guided by Cramer's V identifies eight informative inputs, including aroundhead, backhand, landing height, and player/opponent area coordinates, with a 5-fold CV framework and 300 training epochs. The approach culminates in a runner-up finish at IJCAI CoachAI Badminton Challenge Track 2 and is supported by publicly available code, highlighting practical potential for AI-assisted coaching and strategic planning in turn-based sports, while suggesting future work on correlation-aware embedding strategies.

Abstract

The increasing use of artificial intelligence (AI) technology in turn-based sports, such as badminton, has sparked significant interest in evaluating strategies through the analysis of match video data. Predicting future shots based on past ones plays a vital role in coaching and strategic planning. In this study, we present a Multi-Layer Multi-Input Transformer Network (MuLMINet) that leverages professional badminton player match data to accurately predict future shot types and area coordinates. Our approach resulted in achieving the runner-up (2nd place) in the IJCAI CoachAI Badminton Challenge 2023, Track 2. To facilitate further research, we have made our code publicly accessible online, contributing to the broader research community's knowledge and advancements in the field of AI-assisted sports analysis.
Paper Structure (11 sections, 3 equations, 3 figures, 2 tables)

This paper contains 11 sections, 3 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Evaluation Metrics of IJCAI CoachAI Badminton Challenge Track 2
  4. Dataset and Data Preprocessing
  5. Network Architecture Design
  6. Loss Function Design and Hyper-parameter tuning
  7. Results
  8. Loss Selection Module
  9. Optimal Parameter for MuLMINet
  10. Discussions
  11. Conclusions

Figures (3)

  • Figure 1: Cramer's V Correlation Matrix between features: PL denotes player, AH denotes aroundhead, BH denotes backhand, LH denotes landing height, LA denotes landing area, PLA denotes player location area, and OLA denotes opponent location area
  • Figure 2: Network Architecture: MuLMINet
  • Figure 3: Loss Selection Module