A neurocontroller based on model feedback and the adaptive heuristic critic

A neurocontroller is presented which seeks to minimize a reinforcement signal presented to the controller over time. The architecture is based on P.J. Werbos's (1989) backpropagated critic and incorporates A.G. Barto and R.S. Sutton's (1983) adaptive heuristic critic. Preliminary experiments performed with the cart-pole problem indicate that the method is faster, yet more unstable than a comparable method. The performance is significantly enhanced when a simple proportional control loop is used in conjunction with the neurocontroller