New game engines are heavily based on Computational Intelligence as this can provide game variability and adaptation to the player. This is particularly true for the rehabilitation domain, in which games are used to guide the patient, at home, through the rehabilitation exercises prescribed in the hospital. We describe here our approach implemented in the Intelligent Game Engine for Rehabilitation (IGER). Bayesian adaptation is used to adapt on-line the game parameters to provide an adequate challenge level to the patient looking at the patient's actual performance in the game. A fuzzy system has been implemented to use the knowledge by the therapists to monitor in real-time how the patient is doing the exercises, advising him and avoiding maladaptation. We show also some results on stochastic learning automata in modeling game interaction and how results on graph theory can be used to predict the time course of interaction. Examples in both domains will be shown and discussed.