We examine the problem of a sensor communicating over a wireless channel to an actuator in order to control a plant that is perturbed by a random disturbance. By allowing the sensor to adapt online to the stochastic system state, we develop transmission policies with guarantees on average control performance and required average communication resources. More specifically we design policies with guarantees either on a linear combination of these two objectives, or with guarantees with respect to a hard constraint on the average communication resources. Based on approximate dynamic programming we prove as well as illustrate in simulations that our policies outperform policies that do not adapt online to the stochastic system state.