Modeling Human Spatial Navigation Using A Degraded Ideal Navigator

Navigation of novel environments requires an ability to acquire, store, and recall spatial information. A navigator can use this information to generate novel, efficient paths between arbitrarily-chosen locations in the environment. In the first section of this thesis I review relevant recent research into exploration and spatial learning. I examine biologically-inspired computational models of those processes, and examine the use of virtual reality experiments as a way to test such models. In the second section I present materials, methods, and results for two virtual reality experiments. In these experiments, subjects play virtual taxicab drivers, delivering passengers to stores in a virtual town. In Experiment One, subjects navigate a small town and learn to generate optimal paths after two or three deliveries. In Experiment Two, the size of the town is increased, and a corresponding drop in performance is observed. The third and largest section of this thesis presents a model aimed at exploring the general mechanisms involved in learning the virtual towns. I propose a model whereby (a) subjects build a cognitive map of their environment which becomes more detailed with experience, and (b) they use information stored in the cognitive map to generate delivery paths between arbitrary points in their environment. The model is based on an ideal navigator, which predicts performance of the perfect subject. For Experiment One, the ideal navigator model successfully predicts the delivery path distance for between 60 and 70 percent of the observed deliveries. For Experiment Two, the ideal navigator predicts between 40 and 50 percent of the delivery path distances observed. By systematically degrading the abilities of the ideal navigator, I account for the drop in performance in the larger environment. After parameter tuning, the performance of the degraded ideal navigator model fits mean learning curves for both experiments (using a different set of parameters for each experiment). In addition, the model is able to cluster environments in each experiment based on an estimated difficulty rating. MODELING HUMAN SPATIAL NAVIGATION