Abstract
Shifts in spatial patterns produced during the execution of a navigational task can be used to track the effects of the accumulation of knowledge and the acquisition of structured information about the environment. Here, we provide a quantitative analysis of mice behavior while performing a novel goal localization task in a large, modular arena, the HexMaze. To demonstrate the effects of different forms of previous knowledge we first obtain a precise statistical characterization of animals' paths with sub-trial resolution and over different phases of learning. The emergence of a flexible representation of the task is accompanied by a progressive improvement of performance, mediated by multiple, multiplexed time scales. We then use a generative mathematical model of the animal behavior to isolate the specific contributions to the final navigational strategy. We find that animal behavior can be accurately reproduced by the combined effect of a goal-oriented component, becoming stronger with the progression of learning, and of a random walk component, producing choices unrelated to the task and only partially weakened in time.