A learning automaton approach to trajectory learning and control system design using dynamic recurrent neural networks

A new approach based on the theory of learning automata is presented for training neural networks with recurrent connections and dynamical processing elements. This approach does not require gradient computations and hence affords a simple implementation. Both linear and nonlinear reinforcement actions are suggested which result in specific training algorithms. Applications of the method to two specific problems, viz. learning of continuous-time trajectories and control system design to stabilize a nonlinear dynamical plant, are outlined