THE SELECTION OF MEASUREMENTS FOR PREDICTION.
Abstract
This research is concerned with prediction and pattern recognition. The purpose is to study relationships between a machine's performance and the number and quality of its measurements and to devise techniques of measurement selection and processing which yield minimum error. The major emphasis is not prediction, because it provides the more tractable model for analysis. The primary effort is directed toward the development of prediction or pattern-recognition techniques applicable to problems involving relatively many measurements and a limited number of learning samples. Consideration of the normal prediction model leads to particular methods of measurement selection and processing. The more successful of these techniques reduce the dimensionality seen by the processor, either by selecting measurements or by linearly combining them to form a smaller set of new measurements. Experimental evidence suggests that these special processors, when properly applied, perform substantially better than do conventional linear methods. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Nov 01, 1964
- Accession Number
- AD0456770
Entities
People
- D. C. Allais
Organizations
- Stanford University