The question addressed in this work is 'What do people exactly typically do, if they interact strategically in games they have not much experience with?'. It is certain that human behavior in strategic interactions and games deviates from predictions of game theory. But, it is also certain that this behavior must have some kind of explanation. Eventually, people do not behave in a fully unpredictable way. This work considers general strategic interactions with untrained subjects. It does not consider human performance in well-known games like chess or poker. A very basic scenario is used to investigate human behavior. This scenario is a repeated zero sum game with imperfect information. An experiment with subjects is conducted and the data is analyzed using a set of different machine learning algorithms. As the result, a way of using machine learning is given. Finally, designing a formalism for representing human behavior is discussed.