A Systematic Analysis of Unsupervised, Supervised, Few-Shot, and Neural Networks in Understanding Player Skill in Video Games

Abstract

The video game industry has grown massively. With leaderboard and player statistics becoming readily available, understanding player skill levels can aid in making market strategies, promotions, and advertisements decisions. This research aims to supply a systematic examination of how effective unsupervised, supervised, combined, few-shot, and neural networks (NN) are in automatically deciding player levels in video games. We focus on first-person shooter games such as Call of Duty (COD) and Tom Clancy’s Rainbow Six Siege (RSS). In addition to a systematic performance comparison, we aim to understand methods to reduce the human labeling effort in the work by using two sets of player data. Neural network frameworks such as TensorFlow are viewed, and feature selection of machine learning techniques is called for.