Genetically adaptive neural network classification systems and methods

• US 20050049983A1
• Filed: 08/29/2003
• Published: 03/03/2005
• Est. Priority Date: 08/29/2003
• Status: Active Grant

• Current Assignee: Raytheon BBN Technlogies Corp. (Rtx Corporation)
• Original Assignee: BBN Technologies (Rtx Corporation)
• Inventors: Stickels, Ernest Scott, Butler, Gary Dana, Kanyuck, Robert Warren, Coon, Andrew Charles
• Granted Patent: US 7,324,979 B2
• Time in Patent Office: Days
• US Class Current: 706/13
• CPC Class Codes: G06F 2218/12 Classification; Matching
• CPC Class Codes: G06N 3/086 using evolutionary algorithms, e.g. genetic algorithms or genetic programming

Abstract

Genetically adaptive neural network systems and methods provide environmentally adaptable classification algorithms for use, among other things, in multi-static active sonar classification. Classification training occurs in-situ with data acquired at the onset of data collection to improve the classification of sonar energy detections in difficult littoral environments. Accordingly, in-situ training sets are developed while the training process is supervised and refined. Candidate weights vectors evolve through genetic-based search procedures, and the fitness of candidate weight vectors is evaluated. Feature vectors of interest may be classified using multiple neural networks and statistical averaging techniques to provide accurate and reliable signal classification.

Claims

1. A system for genetically adaptive signal classification comprising:

a learning supervisor for processing a population of weight vectors in a neural network using training data;

a fitness evaluator coupled to the learning supervisor for evaluating a fitness of the weight vectors against a stopping criterion based on the output of the weight vector processing by the learning supervisor; and

a genetic operator, coupled to the fitness evaluator and the learning supervisor, for modifying the population of weight vectors and providing the modified weight vectors to the learning supervisor until the fitness evaluator indicates that the fitness of the weight vectors meets the stopping criterion.

2. The system of claim 1, wherein the learning supervisor has an input port to receive an initial set of random weight vectors.

3. The system of claim 1, wherein the genetic operator includes a selection operator for selecting weight vectors proportional to their fitness and for discarding weight vectors not meeting a specified level of fitness.

4. The system of claim 3, wherein the genetic operator further includes a reproduction operator for duplicating the selected weight vectors to produce a population with the same number of weight vectors prior to selection by the selection operator.

5. The system of claim 4, wherein the genetic operator further includes a crossover operator for exchanging weight vector information around a crossover point.

6. The system of claim 5, wherein the crossover operator randomly selects a variable number of weight vector pairs from a current population before exchanging weight vector information around the crossover point.

7. The system of claim 6, wherein the genetic operator includes a mutation operator for randomly changing weight vectors.

8. A method for genetically adaptive signal classification comprising:

processing a population of weight vectors in a neural network using training data;

evaluating a fitness of the weight vectors against a stopping criterion based on the output of the weight vector processing; and

modifying the population of weight vectors and evaluating the fitness of the modified weight vectors until the fitness meets the stopping criterion.

9. The method of claim 8, wherein processing the population of weight vectors includes receiving an initial set of random weight vectors.

10. The method of claim 8, wherein genetically modifying the population of weight vectors includes selecting weight vectors proportional to their fitness and discarding weight vectors not meeting a specified level of fitness.

11. The method of claim 10, wherein genetically modifying the population of weight vectors further includes duplicating the selected weight vectors to produce a population with the same number of weight vectors as the initial population.

12. The method of claim 11, wherein genetically modifying the population of weight vectors includes exchanging weight vector information around a crossover point.

13. The method of claim 12, wherein exchanging weight vector information includes randomly selecting a variable number of weight vector pairs from the current population before exchanging weight vector information around the crossover point.

14. The method of claim 13, wherein genetically modifying the population of weight vectors includes randomly changing weight vectors.

15. A system for genetically adaptive signal classification of sonar waveforms comprising:

a learning supervisor for processing a population of weight vectors in a neural network using training data;

a fitness evaluator coupled to the learning supervisor for evaluating a fitness of the weight vectors against a stopping criterion based on the output of the weight vector processing by the learning supervisor; and

a genetic operator, coupled to the fitness evaluator and the learning supervisor, for modifying the population of weight vectors and providing the modified weight vectors to the learning supervisor until the fitness evaluator indicates that the fitness of the weight vectors meets the stopping criterion.

16. The system of claim 15, further comprising:

mission data collected in real time using sonar waveforms; and

a classifier for using the modified weight vectors to classify the mission data using one or more neural networks, the classifier classifying whether the input waveforms are targets or clutter.

17. The system of claim 16, wherein the learning supervisor has an input port to receive an initial set of random weight vectors.

18. The system of claim 16, wherein the genetic operator includes a selection operator for selecting weight vectors proportional to their fitness and for discarding weight vectors not meeting a specified level of fitness.

19. The system of claim 18, wherein the genetic operator further includes a reproduction operator for duplicating the selected weight vectors to produce a population with the same number of weight vectors prior to selection by the selection operator.

20. The system of claim 19, wherein the genetic operator further includes a crossover operator for exchanging weight vector information around a crossover point.

21. The system of claim 20, wherein the crossover operator randomly selects a variable number of weight vector pairs from a current population before exchanging weight vector information around the crossover point.

22. The system of claim 21, wherein the genetic operator includes a mutation operator for randomly changing weight vectors.

23. The system of claim 22, wherein the classifier uses a subset of the fittest weight vectors to classify the mission data.