A Comparison of Feed-Forward Networks and Maximum Likelihood on a Point- Source Location Problem
Abstract
The problem of point source location using a multi-beam focal-plane staring array radar is addressed. It is viewed as one in functional approximation in which the position of the source is regarded as a nonlinear function of the sampled radar image and it is required to construct an approximant, using a training set, which minimises the mean square error in the position estimate. The problem is also one of generalisation, since the expected operating conditions are likely to be corrupted by noise and this must be taken into account when designing the approximant. Two feed-forward network architectures are considered - a particular radial basis function network which arises as a consequence of the minimum mean square error solution and is appropriate when the signal-to-noise ratio is 'small' and a multi-layer perceptron, chosen for high signal-to-noise ratio approximation. The errors in the position estimates for each of these approaches are compared with a maximum likelihood position estimation method. The maximum likelihood method gives better overall performance and has the advantage that it is not dependent on the signal-to-noise ratio.
Document Details
- Document Type
- Technical Report
- Publication Date
- Apr 01, 1991
- Accession Number
- ADA247362
Entities
People
- A. R. Webb
Organizations
- Royal Signals and Radar Establishment