Sciences des données pour un jeu vidéo

Details

Supervisors

Delvenne, Jean-Charles

Degree label

Master [120] : ingénieur civil en science des données, à finalité spécialisée

Abstract

enLeague of Legends is a competitive game with thousands of players. Its global competition breaks audience records every year in the world of video games. Therefore, all players dream of becoming better, and in particular, of knowing what really influences the outcome of a game at their level, and whether it is possible to predict the outcome of a game in advance, based on the history of each player and the champions chosen for the current game. This is essentially why this work was created. In order to answer this important question of predicting game outcomes, we identified a number of characteristics related to victory. We chose a sample of 200 representative players and their last 100 games. Then, we tested 5 major machine learning models: Logistic regression, K- Nearest Neighbors, DecisionTree, RandomForest, and support vector machine. These allowed us to understand whether it was possible to predict the outcome of a game even before it had been played. By running the models once over our players, and once for each division, we found that it was generally better to predict game results based only on games of the same level. This highlighted the differences between levels of play and the behaviors of the players at each level. Furthermore, we found that for all levels of players, the character trait having the greatest importance is the win rate of the player on the selected champion for the game. The second most frequently appearing characteristic is the win rate on the red side of the map, which corresponds to the top of the game map. Other traits have varying and lesser importance for each level of play. We believe that it is possible to refine these results by taking into account the roles played by each player. In conclusion, it is possible to predict the outcome of a League of Legends game, even while it is still loading, and achieve levels of precision of around 68%. However, it would be possible to obtain even more precise results by creating models that specifically target each possible role in the game.