TL;DR
This paper introduces a novel approach to predict the outcome of League of Legends encounters using physiological sensor data, moving beyond traditional in-game metrics. Their Transformer model achieves promising results in predicting player performance 10 seconds in advance, paving the way for a potential burnout detection system.
What Problem Does This Paper Address?
The burgeoning field of eSports, with its professional leagues and significant revenue, currently lacks robust tools for player training and performance analytics. Existing methods primarily rely on in-game data (like kills, deaths, and assists), which are prone to becoming obsolete due to frequent game updates. Furthermore, these methods often overlook the player's physical and mental state, which can significantly impact performance. This paper aims to bridge this gap by leveraging physiological sensor data to predict in-game encounter outcomes, offering a more robust and potentially burnout-aware analytics solution.
Methods
The researchers developed a comprehensive sensing system to collect a wide array of physiological and environmental data. This system included sensors for:
- Electromyography (EMG): Measuring muscle activity in the arms, correlating with mouse clicks and keystrokes.
- Inertial Measurement Units (IMUs): Tracking hand, chair seat, and chair back movements.
- Galvanic Skin Response (GSR): Indicating arousal and stress levels.
- Pulse Oximeter: Measuring heart rate.
- Eye Tracker: Recording saccades and eye movement.
- Facial Skin Temperature: Captured by an infrared camera, related to mental load.
- EEG Headset: Measuring brain activity.
- Environmental Sensors: Monitoring temperature, humidity, and CO2 levels.
- Input Logger: Tracking keyboard and mouse activity.
This sensor data was then preprocessed, including resampling to a common time step, outlier clipping, and smoothing. In-game data was used to identify and annotate "encounters" – periods where two or more heroes engage in combat. A player's impact on an encounter was calculated as a cumulative sum of positive (kills, assists) and negative (deaths) events, normalized by the number of participants. This impact was then binarized to represent a "good" or "bad" performance.
Machine learning models, including Logistic Regression, KNN, Gated Recurrent Unit (GRU), and a Transformer network, were trained to predict the binarized encounter outcome up to 10 seconds in the future using the preprocessed sensor data.
Datasets
The study utilized data collected from 10 participants playing League of Legends over three experimental days. Data from the first two days was used for training, and the third day for testing. A total of 888 encounters were analyzed, with 56.1% classified as positive outcomes and 43.9% as negative. The paper also makes this dataset available for public use.
Key Insights and Ideas
- Physiological Data is Predictive: The study demonstrates that physiological sensor data, when analyzed with advanced machine learning models, can predict in-game outcomes with significant accuracy, outperforming models that rely solely on historical in-game data.
- Transformer Architecture Excels: The Transformer neural network architecture proved to be the most effective for this task, likely due to its ability to capture complex temporal dependencies in the sensor data.
- Burnout Detection Potential: The developed model shows promise as a burnout or fatigue detector. By predicting a player's likelihood of losing an upcoming encounter, it could advise players to take breaks or adopt more defensive strategies.
- Feature Importance Reveals Key Indicators: Analysis of feature importance highlighted specific physiological signals that strongly correlate with performance. For instance, gamma wave intensity from EEG and mouse activity were linked to winning, while vertical chair movements were associated with losing.
- Robustness to Game Updates: Sensor-based models are inherently more robust to in-game meta changes and patches compared to models trained solely on in-game data, offering a more sustainable solution for performance analytics.
Key Results
- The Transformer model achieved a ROC AUC score of 0.706 for predicting whether a player would win an encounter 10 seconds in the future.
- When framed as a burnout detection system, the model could predict a player's loss in the next encounter with 73.5% precision and 88.3% recall for a 10-second forecasting horizon.
- Feature importance analysis identified T7 electrode gamma wave intensity as a strong indicator of good in-game activity, while vertical chair movement (linaccel_z_chair_seat) was a key indicator of an impending loss.
- Environmental factors like higher temperature, humidity, and CO2 levels were found to negatively impact player performance.
Limitations and Future Work
The authors acknowledge several limitations:
- Small Participant Pool: The study involved a relatively small number of participants (10).
- Lack of Human Annotation: The dataset lacks human-labeled annotations for in-game replays, which could provide richer context.
- Limited Game Scope: Data was collected only for League of Legends.
Future work suggested includes:
- Larger-Scale Data Collection: Expanding the dataset with more subjects and diverse games.
- Professional Annotation: Incorporating annotations from professional players and coaches.
- More Complex Models: Exploring more advanced machine learning models.
- Deeper Understanding of Physiological Behavior: Further investigating the relationship between physiological signals and player performance.
Summary updated