Immunized Neural Network for Sensory System Fusion

NETROLOGIC proposes to build a system which has the ability to self-modify its own architecture such - that it will learn to fuse data from disparate sources into a smoothly integrated information base containing all - the information that the system was able to extract regarding the features to enable neuro-muscular control commands. Our innovation, application of the immune neural network, utilizes principles of the biological immune system and genetic algoritms. The proposed ontogenic system is analogous to the immune system in that it copies the immune system design of having a constant and variable portion of the antigen. The first part is a constant neural network that is trained initially from what is known about the system to arrive at basic neural network components similar to biological gene structures. Once the components are identified, the system is frozen and a variable part of the system is trained. Genetic algorithms operate on the neural network building blocks to create more fit combinations of these components. The evaluation function determines which members of a new generation will be allowed to survive and replicate. In a relatively short time, the GA composes a neural network architecture which is quite good at solving the mapping problem it is designed to solve. Our application uses the INN system to learn to fuse data from disparate sensor sources into a unified information system for control of NASA spacecraft subsystems.