tensorflow there is a model test data in a more robust way to get --- moving average model.
Moving average model is defined as: moving average (exponential moving average), otherwise known as an exponentially weighted average (exponentially weighted moving average), may be used to estimate the mean local variables, so that the variable updating history related values over time.
According to my shallow understanding of the role is the moving average model, changes the model parameters jitter smaller, reduce the impact of noise on the model, the model is more stable.