Learning processes are often analyzed using a model with a structure that is assumed to apply to all points in time, but with one or more parameters that may change over time. In some applications, such as learning theory, the goal is to model the character of change itself; observations at all time points are used to estimate, for example, learning curves. In other applications, such as tutoring systems, the goal is to make decisions based on status at each point in time. Current observations depend directly on current value of the parameter(s) of interest, but recent observations were produced by values that were probably similar if not identical. This note describes and illustrates approximations for the current value of such a parameter, using models that posit no change but downweight the influence of observations as they recede in time. The idea is similar to that of robust estimation procedures that downweight the influence of outliers. The applicability of these approaches to intelligent tutoring systems (ITS's) that use Bayesian inference networks to update student models is discussed.