J.T. Smith (USA)
online learning neural networks, verification, validation, human factors 1.0
This paper presents the results of research related to the verification and validation (V&V) of on-line learning neural networks (OLNN) that continually adapt and evolve. Such is the situation with OLNN-based air-flight controllers developed at the Institute for Scientific Research, Inc. (ISR). For such systems to be employed in commercial aircraft, a certification process is required that addresses the capabilities and risks associated with a system that is continually adapting and evolving. This research is addressed by the NASA funded project "Development of Methodologies for Independent Verification and Validation of Neural Networks," Research Grant NAG5 12069. ISR has devoted much effort to understanding the complexity of this problem and to the development of approaches to address this V&V requirement.1 Because of adaptation with use, an OLNN cannot be pre-certified at release time, simply based on the analysis of initial training sets.2 The findings and approaches presented here are based upon an analysis of the foundation principles and techniques that underpin the pilot certification process by which pilots are deemed sufficiently prepared to operate an aircraft. Such areas as human factors analysis and accident analysis have been extrapolated to identify corresponding problems and opportunities in addressing the certification of an OLNN-based system.
Important Links:
Go Back
