FUNCTION MODELING EXPERIMENTS.
Abstract
The results of an experimental investigation of the capabilities and the limitations of trainable machines for use in function modeling have been presented. In the simpler experiments the data used for training were noise-free. Here the structure of the machine set the fundamental limitation on the possible performance. In the prediction experiments, random fluctuations in the data also limited performance. Modified training rules, which have not yet been theoretically justified, were developed to cope with the second limitation. Elementary functions were approximated in many of the function modeling experiments; for these functions, the use of a learning machine instead of more conventional methods cannot be justified. However, for more difficult applications, such as those encountered in the prediction experiments, the machine performance was sufficiently good to make the speed advantages of a training machine a significant consideration. Viewed as a statistical procedure, this technique is essentially a non-parametric procedure that can be efficiently performed by a special-purpose computer of the type currently being constructed for pattern recognition problems. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 01, 1963
- Accession Number
- AD0427018
Entities
People
- J. W. Machanik
- R. O. Duda
Organizations
- SRI International