| PROPOSAL NUMBER: | 03-T1.01-9874 (For NASA Use Only - Chron: 030126) |
| RESEARCH SUBTOPIC TITLE: | Information Technologies for System Health Management, Autonomy and Scientific Exploration |
| PROPOSAL TITLE: | A Formal Method for Verification and Validation of Neural Network High Assurance Systems |
| SMALL BUSINESS CONCERN (SBC): | RESEARCH INSTITUTION (RI): | ||
| NAME: | PROLOGIC, INC | NAME: | Institue for Scientific Research, Inc. |
|---|---|---|---|
| ADDRESS: | 1000 Technology Drive | ADDRESS: | 320 Adams Street, PO Box 2720 |
| CITY: | Fairmont | CITY: | Fairmont |
| STATE/ZIP: | WV 26554-8824 | STATE/ZIP: | WV 26555-2720 |
| PHONE: | (304) 363-1157 | PHONE: | (304) 368-9300 |
PRINCIPAL INVESTIGATOR/PROJECT MANAGER
(Name,Email)
Marjorie A. Darrah
mdarrah@isr.us
U.S. Citizen or Legal Resident: Yes
TECHNICAL ABSTRACT (LIMIT 200 WORDS)
Our proposed innovation is to develop neural network (NN) rule extraction technology to a level where it can be incorporated into a software tool, we are calling NNRules, which captures a trained neural network?s decision logic and uses it as a basis for verification and validation (V&V) of the neural network. This formalism has never been attempted. The significance of the NNRules innovation is that:
? The National Aeronautics and Space Administration, the Department of Defense, the Department of Energy, and the Federal Aviation Administration are currently researching the potential of neural networks in mission- and safety-critical systems.
? High assurance neural network applications require rigorous verification and validation techniques.
? The adaptive and ?black box? characteristics of neural networks make verification and validation of neural networks practically intractable.
? Rule-based systems have a more visible, and potentially human readable, decision logic that supports a robust set of verification techniques.
? Neural network rule extraction research has developed algorithms that translate a neural network into an equivalent set of rules. NNRules embeds this technology in a generally usable tool that will dramatically increase the ability to V&V high assurance neural networks.
POTENTIAL NASA COMMERCIAL APPLICATIONS (LIMIT 150 WORDS)
Possible NASA neural network applications include adaptive flight control, mission planning, deep space autonomous operations, and vehicle health monitoring. Our proposed innovation provides R&D of V&V methods for neural networks, potentially assisting in the increased use of neural networks. Adaptive flight control neural networks can correctly control the vehicle in response to unknown or unforeseeable situations. In deep space, the closed loop command path between spacecraft and Earth is too long in duration to ensure ground based response to critical events. Neural Nets enable the spacecraft to make critical decisions autonomously to both maintain system health and enhance science collection.
POTENTIAL NON-NASA COMMERCIAL APPLICATIONS (LIMIT 150 WORDS)
The Navy has been researching incorporating neural networks into ship operations as an enhancement to its Smart Ship technology. With neural networks, the trust in ship critical operations is compromised until they can be verified and validated. FAA interest in neural networks sparked a joint effort with NASA ARC, researching how neural networks can be used in flight control systems and successfully testing neural networks in critical systems on the DFRC F-15 test plane. Intelligent planes with neural networks can learn permissible flight envelopes and be trained not to fly outside those envelopes, circumventing a repeat of 9/11.