Generation and testing of hypotheses are central aspects of the process of scientific discovery. In instructional contexts, students can acquire a basic understanding of these concepts by means of scientific discovery learning. In order to study the mechanisms that underly students' ability to generalize from specific observations and to use these hypotheses to derive predictions, observations on students working in a computerized discovery learning environment for geometrical optics are analyzed. In addition, cognitive simulation programs taking the form of production systems are developed to capture the central aspects of students' discovery learning strategies. These task-specific models are discussed within the framework of general computational theories of human inductive learning.