On using Artificial Neural Network models to predict game outcomes in Dota 2

Sammanfattning: Dota 2 is an online strategy game, played in a five versus five format. Its multitude of selectable characters, each with a unique set of abilities and spells, causes every new match to be different from the last and picking the right characters can ultimately decide whether a team wins or loses a game. This report investigates if Artificial Neural Networks can be used to predict game outcomes, based solely on the character selection made in each game. Additionally, the report will explore if altering the base parameters of the utilized ANN models can improve predictive performance. The models considered in the thesis will thus vary in number of hidden neurons and hidden layers in the neural network. The results show that the various models have an average prediction accuracy ranging between 53-59%. We find that using a low number of neurons with many layers improves prediction accuracy. Although the results from this study seem to indicate a correlation between character picks and game outcomes, we recommend a more extensive analysis be conducted in order to reproduce these results and thus ensure external validity.