This work concerns the control of stochastic systems with unknown and randomly time-varying parameters. Cost functions which consider the sum of output variances up to M steps ahead in time are adopted in the optimization of the control performance. Optimal predictors are used to replace the future outputs which are needed in the solution of the optimization problem. The consequences of this simplification are investigated. A formula is obtained for the computation of the control signal in the rase of M-steps-ahead optimization. The relationships between the controller presented here and other classical suboptimal dual controllers are analyzed. Simulation results illustrate the actual performance of the new controller.