Design Space Exploration for Cyber Physical Systems
Abstract
The goal of this effort was to evaluate the feasibility of creating a toolkit for design-space exploration of cyber physical systems and identify toolkit success metrics. This toolkit needed to accomplish tasks, including: synthesis or inverse design, composition of heterogeneous models, alleviating curse of dimensionality, minimizing blackbox function evaluation, resilience to blackbox evaluation failure, non-parametric synthesis, solution to nonlinear constraints and optimization. This research proposed Constraint satisfaction with Neural networks, Mixed integar linear programming (MILP) and Active learning (CNMA), a new method of solving nonlinear constraints on nonlinear functions. CNMA is based on the idea that the constraints can be solved by learning the function as a neural network, converting it into an equivalent MILP, and solving with industrial-strength MILP solvers. Since learning is always approximate, an incorrect solution can be returned. CNMA adds a new error-correction step that assures solution correctness. Additionally, it focuses learning of the function only on that part of its domain that is relevant to solving the constraint. Thus, the learning is reduced by orders of magnitude over the case where the function has to be learnt in its entire domain.
Document Details
- Document Type
- Technical Report
- Publication Date
- Nov 01, 2019
- Accession Number
- AD1086534
Entities
People
- Brendan Englot
- Emily Mak
- Jeremy M. Cohen
- Kishore Pochiraju
- Niraf Jha
- Sanjai Narain
- Todd Huster