Protocols were obtained from 22 subjects as they discovered the conditions under which equilibrium is obtained on a balance beam by predicting and observing the outcomes of a series of problems. The interviews revealed that subjects used a variety of heuristics to make predictions once they had isolated the two relevant features of the problem, weight and distance, but before they could appropriately relate them in one procedure. These heuristics included instance-based reasoning, qualitative estimation of distance, and the use of quantitative rules of limited generality. Variability in the use of these heuristics by individuals suggests that learning to understand the balance beam is not best described in terms of a simple rule acquisition theory, and that reasoning from instances plays a more important role in inferring a physical relationship than previously thought. (Author)