An abstract artificial neural architecture for efficient, reliable, and adjustable solutions to difficult learning control problems

A new artificial neural architecture designed for efficient, reliable, and adjustable solutions to difficult learning control problems is introduced. It consists of a hierarchy of command and control centers which govern motor selection networks. Internal drives, similar to those in biological systems, are formed within the controller to facilitate learning. Efficiency, reliability and adjustability of this architecture are demonstrated on the benchmark inverted pendulum dynamic control problem. A comparison with results from artificial learning systems discussed in the literature is given. It is shown that the command and control center/internal drive architecture learns over 100 times faster than Barto, Sutton, and Anderson's (1983) adaptive search element/adaptive critic element system, experiencing less failures by more than an order of magnitude. The preliminary results reported here indicate that the new architecture should be able to handle much larger regulator problems.