Predicting human behavior in smart environments: theory and application to gaze prediction

Predicting human behavior is desirable in many application scenarios in smart environments. Gaze represents one of the essential cues, which is important to understand these behaviors. In this thesis, we consider eye movements and the spatial location of visual attention in different behavioral context as a model system. Behavioral eye movements data in a different context is presented together with predictive models of visual saliency. The existing models for eye movements do not take contextual factors into account. This addressed using a systematic machine-learning approach, where user profiles for eye movements behaviors are learned from data. Machine learning models and the analysis of behavioral data show the limitations of current predictive models describing human eye movements behaviors and reveal the influences of task on gaze selection. The analysis furthermore demonstrates the relative importance given to the individual visual features, and it shows that simple predictive "one-fitsall"-models will not work for eye movements prediction. This part of the work used model-based systematic data analysis. Human studies have shown that eye movements behaviors are mostly effected by the task at hand. For that, human vision has to learn how to move the eyes to the relevant information. In this part of the work a theoretical innovation is presented, which goes beyond pure data analysis. The thesis proposed the modeling of eye movements as a Markov Decision Processes (MDPs). Then it use Inverse Reinforcement Learning (IRL) paradigm to infer the reward function. The examined IRL approaches used information about the possible eye movement positions. We found that it is possible to automatically extract reward function based on effective features from user eye movement behaviors using IRL. We found that the reward function was able to extract expert behavior information that fulfill to predict eye movements behaviors. By using a new inverse reinforcement learning paradigm that constructs the parameters of the learning model to best match the observed human behavior, the connection between model and empirical data is obtained. The application of this method to the empirical data shows that this model can be used in eye movement predictions, and in human behavior modeling in general.