Concepts are central to human cognition and one important type of concepts can be represented naturally with symbolic rules. The learning of such rule-based concepts from examples relies both on a process of perception, which extracts information from the presented examples, and a process of concept construction, which leads to a rule that matches the given examples and can be applied to categorize new ones. This thesis introduces PATHS, a novel cognitive process model that learns structured, rule-based concepts and takes the active and explorative nature of perception into account. In contrast to existing models, the PATHS model tightly integrates perception and rule construction. The model is applied to a challenging problem domain, the physical Bongard problems, and its performance under different learning conditions is analyzed and compared to that of human solvers.