Stats4Grads: Examining the Effect of Uncertain Preferences upon Value of Sample Information
29 January 2007 13:00 in CM105
The use of the Bayesian paradigm coupled with acceptance of the Expected Utility Hypothesis provides a powerful and philosophically compelling methodology for decision making in situations involving uncertainty. However, it is traditionally assumed that the decision maker is able to correctly state her utility function for any possible reward realisation. The theory of Adaptive Utility addresses this problem, seeking to create a normative decision theory for when utilities, and hence preferences, are uncertain.
Under this setting a DM is permitted to be surprised by obtaining actual utility different to that which was expected under initial beliefs. Such an outcome is an example of how a decision maker may learn about her true preferences following decision selection, and an important distinction from classical theory is that in a sequential setting, the optimal initial decision need no longer correspond to that which would have been determined if it were assumed utilities were fixed at initial expected values.
This talk examines the effect of permitting uncertain utility upon the classical Bayesian decision theory meaning of Value of Information. Not only can a decision maker collect sample information on the underlying processes that determine how results are obtained, but now they may also be presented with opportunity to learn about their correct utility function. The relationship between `classical' value of information and level of uncertainty over preferences is also discussed.
See the Stats4Grads page for more details about this series.