Diversely enumerating system-level architectures

Abstract

Embedded systems are highly constrained, and modern constraint solvers can be used to synthesize architectures satisfying constraints. However, solvers may enumerate solutions with bias yielding a skewed view of the solution space. In this paper we formalize enumerators, enumeration bias, and build empirically diverse enumerators. An ideal diverse enumerator uniformly draws equivalence classes of solutions. We develop an algorithm, called symmetry-directed randomized partitioning, for constructing empirically diverse enumerators from modern unmodified constraint solvers. We show that our diverse enumerator is significantly closer to the ideal compared to baseline and randomized solvers.