|
← Return to agenda
Meeting | ACGS Committee Meeting 133 - Asheville, NC - November 2024 | Agenda Location | 9 SUBCOMMITTEE E – FLIGHT, PROPULSION, AND AUTONOMOUS VEHICLE CONTROL SYSTEMS 9.2 Testing and Verifying Neural Network Control Systems | Title | Testing and Verifying Neural Network Control Systems | Presenter | Doug Wehbe | Affiliation | Stony Brook University | Available Downloads* | presentation | | *Downloads are available to members who are logged in and either Active or attended this meeting. | Abstract | Neural network control systems (NNCS) are closed-loop systems where a neural network controller actuates a plant governed by differential equations. Verification is often challenging to scale for this class of systems, as multi-step analysis requires repeated rounds of both neural network verification and reachability analysis. Recently, a proposed approach improved analysis efficiency by first approximating the neural network component using an approximation based on input quantization. This simplifies neural network analysis, allowing improved scalability for closed-loop verification. In this talk, we present IQ-Verify, a general tool for input-quantized verification that expands the practical applicability of the approach. We demonstrate applying the tool to the closed-loop ACAS Xu air-to-air collision avoidance system. We further show how IQ-Verify can be used to falsify the original, non-quantized systems. | |
Copyright © 2025 | Question? webmaster@acgsc.org
|