Experimentation with heteroskedastic noise (Q1341489)

In models of active learning or experimentation with noisy observations, the informativeness of the signal obviously depends on the (not always evident) relationship between the choice variable and the noise. This previously implicit relationship is made explicit by introducing heteroskedastic noise into the model. The general conditions for the informativeness of the signal to be independent or decrease with respect to the choice variable are obtained. Furthermore, the model reveals that many of the results in previous models of learning do not arise from what is unknown nor the existence of uncertainty per se, but rather from how incomplete information is modeled.