Browsing by Author "Kujala, Janne"

Filter results by typing the first few letters
Now showing 1 - 1 of 1
  • Bayesian student modeling in a learning game
    (2010) Ruohonen, Rauli
     School of Science | Master's thesis
    The possibility of early identification of dyslexia has motivated the development of the computer game "Graphogame", which is designed for the early prevention of reading difficulties. It is important that the learning content in the game is adapted for individual players so that the game will be neither too difficult nor too easy. For this purpose, it is helpful to have a precisely defined statistical model of the player. In this thesis, we present a Bayesian model of the player. The model can be used for the selection of effective game content and for tracking the players' progress. The model is essentially a combination of factor analysis and the logistic model. In the game, the players are presented with choice situations, where the letters of the alphabet are displayed on the screen. The task is to select the letter that corresponds to the sound that is played at the same time. Information about the choice situations and the choices the players make is stored in game logs, which we use to fit the model. We are basically performing a factor analysis on the skills of the players, which are observable only through the choices the players make. There are three pre-existing statistical player models for Graphogame, but in all of them only one player is taken into account at a time, independently of all the other players. In contrast, in the model proposed in this thesis all the players are considered simultaneously, allowing the model to learn statistical properties of the player population. The pre-existing model best suited for the analysis of the game data is used to construct baseline models in this thesis. The employed model and fitting method are tested using simple simulations as well as real game log data. The model works well in the simulations, if there is enough data of each player. The model also predicts player choices in the real data better than the baseline models.