The Spiral Discovery Network (SDN) was recently proposed as a tool for automated parametric optimization based on the Spiral Discovery Method. SDN can be seen as a heuristic optimization approach that offers tradeoffs between exploration and exploitation without hav- ing recourse to explicit gradient-based feedback information (unlike classical neural networks) and without requiring hand-coded representations of metaheuristic constructs such as genotypes (unlike genetic algorithms). In this paper, the properties of the SDN model are further explored, and two extensions to the model are proposed. The first extension corrects a shortcoming of the original model and has to do with the assignment of credit among different output components based on the most recent performance of the model at any given time. The second extension con- sists of using multiple SDN cells in a hierarchical architecture, which enables a fuller and more effective exploration of the parametric space. The improvements provided by the two extensions are validated on the same set of simulations discussed in earlier work.