Solutions to a class of linear-quadratic-Gaussian (LQG) stochastic team problems with nonclassical information (Q911536)

An information structure is called nonclassical if the action of decision maker j affects the information of i and there is no way in which i can infer the information available to j. It is now well-known that, under such information structure, the optimal decisions for LQG problems do not have to be affine functions of the observations. It is shown in the present paper by means of two specific problems that, for some class of LQG team problems, the search for an optimum may still be confined to affine policies. The problems posed may be viewed as the transmission of a noisy version of a Gaussian message over a number of noisy channels under a fidelity criterion. In the case where the channel noise is uncorrelated with the encoder input, the results of the paper show that the optimal solution may be interpreted as first extracting the message from its noisy version under a mean square criterion and then transmitting this extracted message.