Designing the Architecture of Hierachical Neural Networks Model Attention, Learning and Goal-Oriented Behavior
Final rept. 1 Nov 1988-31 Dec 1993
DREXEL UNIV PHILADELPHIA PA DEPT OF ELECTRICAL AND COMPUTER ENGINEERING
Pagination or Media Count:
During this period this grant partially supported 6 researchers, and resulted in over 21 publications. This unusually large activity is largely due to the enthusiasm of the researchers and their institution, Drexel University, which indirectly carried some of the financial burden. Neural or other learning architecture for real world, real time applications, necessarily employ feedback and thus deal with the unavoidable dilemma of identification versus stabilization or tracking. The major finding reported focuses on this tradeoff and how to optimally perform it. For linear time invariant finite dimensional systems they are able to perform on-line closed loop identification and tracking. If in addition the learning and tracking cost functions are quadratic they show these costs may be linearly scalarized without loss of optimality.