Connectionist Models for Intelligent Computation

Abstract

Abstract : We have successfully demonstrated the capability of neural net to learn the generation grammar and automata for discrete symbolic sequences. The first level of complexity is represented by the regular grammar and can be learned by a recurrent neural net. The capacity of these recurrent neural net to represent finite state machines that generate these regular grammar has also been theoretically estimated. In the next level of complexity, neural net was trained to operate a stack memory to recognize context free sequences. Finally, we showed that recurrent net can be constructed with a neural tape to represent a universal Turing machine. For continuous time series, we also showed that neural net can be used to classify curves with different topology, to control chaos system without pre-knowledge of its fixed points and to perform system identifications for real physical systems. For the last task, we have successfully trained a neural net to simulate the flight dynamics of a helicopter UH-60. The system used is a MIMO model of recurrent net. And the helicopter model has six degrees of freedom, the vertical, side and forward speed, the pitch rate, roll rate, yaw rate and four control maneuvers, the lateral longitudinal, directional and the collective controls.