Optimal control requires the definition of a control policy from the behavior of a plant, without the luxury of a desired reference trajectory. In the case of this work the direct method of optimal adaptive control is taken where feedback from the environment has no sign or directional information. Moreover, the case of continuous valued as opposed to binary valued control actions is required. The proposed architecture demonstrates extensive use of hierarchical partitioning of the problem in order to decompose the task into a composition of subtasks. The significance of variance terms in the design of RBFs is emphasized, and the entire network demonstrated on benchmark nonlinear control tasks. In each case, the emphasis is towards the location of robust solutions without recourse to any a priori information (c.f. linearized plant or fuzzy rules).