Titlebook: Neural Network Perception for Mobile Robot Guidance; Dean A. Pomerleau Book 1993 Springer Science+Business Media New York 1993 Navigation.

Notorious

專(zhuān)心

,Training Networks “On-The-Fly”,the assumed impracticality of training networks to perform realistic high level perception tasks, connectionist researchers have frequently restricted their research to toy problems (e.g., the T-C identification task [Rumelhart, Hinton & Williams, 1986, Jacobs et al., 1990]) or to low level tasks (e

木質(zhì)

Training Networks With Structured Noise,. This technique provides sufficient variety to allow a network to drive reliably as long as the important image features, such as lane markers and road boundaries, remain consistent in appearance between training and testing. However there are situations which give networks trained using the on-the

巨頭

corn732

Analysis of Network Representations,rks. While I would dispute the claim that a “black-box” connectionist solution to a difficult problem is entirely useless, I agree that understanding the details of a solution is important. Understanding the processing being performed can point out situations or conditions where the system is likely

figment

分開(kāi)

Input Reconstruction Reliability Estimation, of the novelty could be used to judge the appropriateness of individual networks for the task to be performed. The location and shape of novel features could be employed to identify the unusual components of the input and to choose an appropriate response.

從屬

Other Applications - The SM2,s flexibility. I describe the application of these methods to the task of precisely guiding a robot called the Self Mobile Space Manipulator (SM.), which is designed to walk on the exterior of space station Freedom [Brown, Friedman & Kanade, 1990, Ueno, Xu & Brown, 1990]. While the SM. is a very dif

發(fā)生

思考而得