We are motivated by the tasking problem for UAVs in an adversarial environment. In particular, we consider the problem where, in addition to purely random noise in the observation process, the opponent may be applying deception as a means to cause us to make poor tasking choices. The standard approach would be to apply the feedback-optimal controls for the fully-observed game, to a maximum-likelihood state estimate. We find that such an approach is highly suboptimal. A second approach is through a concept taken from risk-sensitive control. For the third approach, we formulate and solve the problem directly as a partially-observed stochastic game. A chief problem with such a formulation is that the information state for the player with imperfect information is a function over the space of probability distributions (a function over a simplex), and so infinite-dimensional. However, under certain conditions, we find that the information state is finite-dimensional. Computational tractability is greatly enhanced. A simple example is considered, and the three approaches are compared. We find that the third approach is yields the best results (for such a case), although computational complexity may lead to use of the second approach on larger problems.