AIMED: AI-Mediated Exploration of Design: An Experience Report

Abstract

The goal of DARPA's Symbiotic Design of Cyber Physical Systems (SDCPS) program is to develop tools for “correct-by-synthesis” design of cyber physical systems (CPS) and reduce the time from concept to deployment from years to months. Achieving this goal poses several hard challenges. Design spaces are high-dimensional cross-products of discrete and continuous spaces. It can take minutes to hours to evaluate the performance of a design. The human designer's intent is often not concretely articulated. Sometimes designs are not created from scratch but rather by completing or repairing existing ones. This paper outlines how the AIMED system addresses these challenges. AIMED consists of three core technologies. The first is “deformable connector” that eliminates an important type of discreteness from design spaces. Thus, not only is the design space vastly simplified, efficient optimization engines for purely continuous spaces can be used in the search for a design. The second core technology is Inverse Specification, based on inverse reinforcement learning that infers human intent by asking the human a small number of simple preference questions. The third core technology is Gaussian Mixture Models that allows completion and repair of designs and finds not just one but a diversity of solutions. AIMED is illustrated in the context of Unmanned Airborne Vehicles (UAVs) although it was also applied to the design of Unmanned Underwater Vehicles (UUVs). AIMED was used to automatically discover high-scoring, novel UAVs, unencumbered by biases of planarity and symmetry: a UAV with non-coplanar propellers and another with asymmetric wings. We expect our experience will apply to design of other CPS.