Learning to Make Interactive Decisions. A Case Study of "Rock-Paper-Scissors" Game

The present diploma paper is concentrated on the analysis of how people make interactive decisions. This issue is analyzed through the prism of learning models. This type of models is now being more and more used in economics, especially behavioral and experimental. This work is concentrated on the three most famous of them: Roth-Erev (1998) Reinforcement Learning, Fudenberg and Levine (1995) Belief Learning, the Experience Weighted Attraction Model of Camerer and Ho (1998, 1999). These models are used in order to try to describe a person's behavior when the opponent has a strategy that does not focus on choosing one optimal action, but is rather based on some rule. Experiment of a “Rock-Paper-Scissors” game between the participants (humans) and the computer algorithm (robot), which plays with a fixed randomness probability, provides data to test the models. The results showed that the learning models, such as RL, BL and EWA are good at predicting the participants’ behavior only when the program of the computer is set to total randomness. When the program of the computer is more advanced – the randomness is mixed with the counter Pavlov strategy, the models are far from being optimal in describing the learning process of a player.