A neurocontroller for robotic applications

Adaptive control of a hybrid electric vehicle

“…The power of the NAC in this application is due to the fact that minimal a priori knowledge of the plant dynamics is required to implement the NAC control law: its order, the sign (or orthogonal part) of the input mapping, and a "slow varying" assumption; allowing the control law to adapt in real-time to changing loads, wind conditions, and road surfaces. Indeed, it has been our experience in other applications [7], [35] that, in spite of the slow-varying assumption, the NAC has little difficulty in dealing with abrupt changes in the reference command; sharp turns, abrupt accelerations and braking etc. ; where the controller effectively "reinitializes" itself and adapts to a new set of gains.…”

“…Indeed, it has been our experience in other applications [7], [35], that a NAC controller with no prior learning typically needs to be exercised three or four times to learn the system dynamics sufficiently well to compare favorably with a well tuned nonadaptive controller. Of course, one must tradeoff this deficiency with the adaptive controller's ability to deal with changing and unmodeled dynamics.…”

“…Following Seraji [39], however, and based on our previous experience in robotic and flight control [7], [35], we rely on the adaptivity of the controllers to compensate for the "relatively loose coupling" between the four controlled subsystems. This difficulty could be further alleviated by replacing the three single-input-single-output NAC controllers by a single multivariate (three-input-three-output) NAC "chassis" controller.…”

“…On the other hand, the ADP is a "suboptimal" ADP [4], [28], [47] variant where the control law is based on an approximate solution to Bellman's equation, which is periodically updated until it converges to the optimal control law [8], [28], [36]. These controllers have been used in a number of applications over the past few years: to robotic joint control [7] and flight control [8], [35] as well as the present automotive application [34], while the required algorithms have been detailed elsewhere [8], [28], [36], [37], [39]. As such, we simply provide a short overview of the two algorithms in Appendixes A and B.…”

“…In both cases, the control law is based solely on the error between the actual and desired state, without the intermediary of an explicit system identification process, while both algorithms are characterized by a full stability and robustness theory. The NAC is an -order multivariate derivative [7], [35], [37] of a second order single-variate model reference adaptive controller [39], which is designed to track a prescribed trajectory in state space. On the other hand, the ADP is a "suboptimal" ADP [4], [28], [47] variant where the control law is based on an approximate solution to Bellman's equation, which is periodically updated until it converges to the optimal control law [8], [28], [36].…”