Operations Research (Pattern Classification)
Client Problem
In 1992, the Naval Surface Warfare Center solicited proposals for exploratory development of methods for the processing of sonar signals that would enable earlier classification of submarines. This desire was motivated by recent theoretical advances which had been made in identifying the spectral components of the echo return of wide band sonar signals reflecting from submarine-like targets. The navy sought experimental studies to explore methods for evaluating the vibrational patterns of submarine models which could then be used for early classification of submarines.
What We Did
The navy had originally believed that the most promising techniques for addressing this problem would be found in the developing area of neural networks applied to pattern classification. Our experience in operations research led us to conclude that an optimization based approach to pattern classification, with which we were familiar, represented a potentially promising alternative approach. We then proposed and were awarded a Small Business Innovation Research (SBIR) contract for a program of basic research and exploratory development into an improved signal interpretation technique for classifying acoustic signals based on an approach known as linear programming (LP) discriminant analysis.
Results
Our research into the LP approach to discriminant analysis and pattern classification led to an enhanced understanding of the model and demonstrated the efficacy of the approach to pattern classification problems in a variety of areas. We demonstrated that the approach is generally superior to conventional, statistically based classifiers and showed that the approach yields success rates comparable to those of neural network approaches while possessing some superior operating characteristics. The project resulted in the development of proprietary software for pattern classification that we have applied to numerous other classification problems ranging from digital signal processing (seismic first break picking, vibration analysis of underground mining machinery, and speech classification) to business applications (credit evaluation, international portfolio evaluation, and the assessment of factors in new product development).
In 1992, the Naval Surface Warfare Center solicited proposals for exploratory development of methods for the processing of sonar signals that would enable earlier classification of submarines. This desire was motivated by recent theoretical advances which had been made in identifying the spectral components of the echo return of wide band sonar signals reflecting from submarine-like targets. The navy sought experimental studies to explore methods for evaluating the vibrational patterns of submarine models which could then be used for early classification of submarines.
What We Did
The navy had originally believed that the most promising techniques for addressing this problem would be found in the developing area of neural networks applied to pattern classification. Our experience in operations research led us to conclude that an optimization based approach to pattern classification, with which we were familiar, represented a potentially promising alternative approach. We then proposed and were awarded a Small Business Innovation Research (SBIR) contract for a program of basic research and exploratory development into an improved signal interpretation technique for classifying acoustic signals based on an approach known as linear programming (LP) discriminant analysis.
Results
Our research into the LP approach to discriminant analysis and pattern classification led to an enhanced understanding of the model and demonstrated the efficacy of the approach to pattern classification problems in a variety of areas. We demonstrated that the approach is generally superior to conventional, statistically based classifiers and showed that the approach yields success rates comparable to those of neural network approaches while possessing some superior operating characteristics. The project resulted in the development of proprietary software for pattern classification that we have applied to numerous other classification problems ranging from digital signal processing (seismic first break picking, vibration analysis of underground mining machinery, and speech classification) to business applications (credit evaluation, international portfolio evaluation, and the assessment of factors in new product development).