Abstract
This research presents a real-time win-prediction model for Super Street Fighter II Turbo utilizing Long Short-Term Memory (LSTM) networks. By simplifying the game state into a time series of player health percentages, the model achieves high predictive accuracy (AUC > 0.90) as early as the 75% mark of a round. The study demonstrates that LSTMs outperform more complex Transformer architectures on smaller-scale sequential gaming data, offering a viable path for live broadcast integration with minimal latency.
Introduction
Real-time, mid-match forecasting in fast-paced competitive gaming remains a significant challenge due to the rapid evolution of game states and unpredictable human decision-making. Unlike pre-match predictions based on historical statistics, live win probability requires modeling the immediate "momentum" of a match. This work aims to enhance viewer engagement by providing dynamic win probabilities, similar to those used in traditional sports broadcasts like the NFL or NBA.
Methodology
The researchers compared several neural network architectures for time-series forecasting, specifically focusing on Long Short-Term Memory (LSTM) networks. LSTMs are recurrent neural networks designed to retain information over long sequences, making them well-suited for the fluctuating nature of fighting game rounds.
A central innovation is the "Health-as-Proxy" approach. Rather than tracking complex variables like character positions or specific move inputs, the model uses only the percentage of health lost by each player at 5 frames per second. This simplification assumes that health changes effectively encapsulate the underlying strategies and player performance.
Data and Experimental Results
The study utilized a custom dataset derived from 10 professional Super Street Fighter II Turbo tournament videos, comprising over 274,000 frames.
Performance Comparison
The LSTM model significantly outperformed baseline models (KNN, SVM, Random Forest) and even complex Transformer Encoders. While Transformers are state-of-the-art for many tasks, they were found to be prone to overfitting on this specific dataset size. At 75% round progression, the LSTM achieved an ROC-AUC score of approximately 0.94, compared to the Transformer's ~0.89–0.93 and baseline models which hovered near 0.50–0.60.
Temporal Sensitivity
The model's accuracy scales with round progression. While early-game "poking" provides limited signal, the predictive power jumps significantly by the 75% mark. The LSTM's inference time is under 3ms, ensuring it can be integrated into live streams without noticeable delay. Even at 95% progression, the LSTM maintained a near-perfect AUC of 0.98, showing high robustness against last-second interactions.
Conclusion
The use of LSTMs with simplified health-based features provides a robust framework for real-time win prediction in arcade fighting games. Future research will explore data augmentation to improve generalization and the incorporation of pose estimation to track character-specific movements and player fatigue.
Summary updated