Action classification of humanoid soccer robots using machine learning

This paper presents an alternative approach on humanoid soccer robots action classification in order to seize the ball control and better ball possession using machine learning and data mining classification algorithms. Categorizing proper actions regarding to positional and environmental features is a prerequisite for proper acting in robotics. In this paper we present an approach to gather information and extracting useful features out of that information from SimSpark simulation server logs. These gathered data will generate a meaningful multi-class dataset, afterwards data processing and running appropriate data mining algorithms on the dataset and evaluating our experiments are the most important issues in this paper. In order to achieve a model for classifying our multi-class dataset, we applied two well-known applications from the domain of data mining: TANAGRA and WEKA and finally we have visualized our experimental results as far as possible.