Reconfigurable Manipulators are structurally redundant robots that utilize a subset of their joints to perform a specific task optimally. This paper presents a method of finding a task-based optimal configuration for a new type of reconfigurable robot manipulator, called the modular autonomously reconfigurable serial (MARS) manipulator. The reconfiguration optimization treats the joint space of the MARS manipulator as a 12-dimensional smooth configuration manifold. The manifold is discretized and ranked based on a variety of criteria, and then clustered into attractive and repellent regions. The user then specifies which regions are desired in the target configuration, and the manifold is reduced in dimension in order to maximize the number of attractive regions and minimize the number of repellent regions. Six manipulator configurations are synthesized using this approach, and their effectiveness is compared.