Development of Artificial Intelligence-based Model to Predict Esports Players’ Skill Levels

Evaluating eSports players’ skills could be significantly important for game developers by enabling them to make their designed games adaptable to the players’ skill levels. Moreover, an accurate and reliable method that could discover the features discriminating between low-skilled and highly-skilled players could help eSports teams design a more effective program focused on improving the extracted features. Sensor data, and in particular EEG data, have been proved to be effective in classifying the players’ skill levels. However, there is currently a lack of a reliable EEG dataset in the literature. Moreover, methods using sensor data still lack sufficient accuracy due to several reasons, including inefficient data preprocessing, inefficient feature extraction, and uncertainties and inaccuracies in sensor and EEG data. Novel methods to overcome these problems are proposed in this thesis. For this purpose, a dataset containing EEG data of 18 League of Legends (LoL) players is gathered first. To the best of my knowledge, this is the first reliable EEG dataset with the highest sample size compared to the previous studies. In the next step, two novel methods are introduced to classify players’ skill levels using sensor and EEG data, respectively. For the first method, a new preprocessing technique using a time window that extracts the most relevant segments of players’ sensor data in LoL is developed. Symbolic Transfer Entropy (STE) is then employed to compute the STE features between sensors’ data using the extracted segments. These features, which show the connectivity between body parts, are finally used to classify the players' skill levels and calculate the optimum window size. For the second method, a new brain source localisation method, called ReLORETA, is proposed to reconstruct the EEG signals inside the brain, thereby removing uncertainties that naturally exist in EEG data. The reconstructed brain signals are then used to classify the players into two skill levels. The first and second methods were applied to a secondary sensor dataset, and the primary gathered EEG dataset, respectively. The classification results demonstrated a significant improvement compared to previous methods. The most significant features in the classification of the players' skills were the connectivity features between gaze data and keyboard, mouse, and hand movement. Also, there was a significant difference in the temporal, frontal, and occipital lobes between professional and amateur players. Besides eSports players, the proposed methods in this thesis can be also applied to sports people's data to improve the training programs in sports and eSports industries.