Predicting game results has always been popular in different sports fields and with the development of technology. eSports has become a big part of young people's lives with growing trends in players from casual gameplay to professional gameplay. One of the most popular methods of making predictions is by utilizing machine learning methods and algorithms. By applying machine learning models, complex player behavior can be analyzed for predicting individual and long-term match results. As such, to solve this problem of this work, Logistic Regression (LR), Random Forest (RF), and LightGBM(LGBM) are used to predict the possibilities of a team winning of each round of Counter Strike: Global Offensive games.
Counter Strike is a popular first-person shooting game that has a rich community and plenty of accessible data and results. There are a seemingly endless number of features, events, and environmental factors that have an influence or correlation on gameplay and match results. In this thesis work,instead of player skills or team ratings, only the features that represent the current, realtime game situation (e.g., number of alive players) or features related to spatial data are investigated for modeling and analyzing results via machine learning models. Distance is considered to be a direct way of showing position relations, which ultimately, have a large influence on match results. Therefore, different features related to player-to-player distances and player-to-bomb distances were used to represent the positional relations in the games. Two path planning algorithms, A* and Floyd-Warshall, were used to obtain distances so that they are reflective to actual player distances in the matches.
For the three machine learning algorithms and two path planning algorithms, the data are trained one time without hyper-parameter tuning and one time with tuning. Consequently, there are a total of 12 predictions for a single set of data to compare with the best accuracy's being reported as: 88.1% for LGBM, 87.9% for RF, and 85.9% for LR and an observation that LGBM is more sensitive to parameter tuning. This thesis verifies that the round winning situation can be successfully predicted with satisfactory results. Furthermore, ensemble machine leaning methods work better than Logistic Regression for this specific problem and setup. Lastly, in this work, the Floyd-Warshall algorithm was shown to work better with ensemble methods while A* is better with Logistic Regression.
- JetBrains GoLand 2019.3
- PyCharm Community Edition 2019.3.2
- QGIS3
- LightGBM (https://lightgbm.readthedocs.io/en/latest/)
- NumPy (https://numpy.org/)
- Pandas (https://pandas.pydata.org/)
- scikit-learn (https://scikit-learn.org/stable/)
- SciPy (https://scipy.org/scipylib/
- https://github.com/mholt/archiver
- https://github.com/markus-wa/demoinfocs-golang
- https://github.com/AtsushiSakai/PythonRobotics/tree/master/PathPlanning