Learning capabilities for improving automatic transmission control

We analyzed the gear-box position selection (GPS) problem on automatic transmission (AT) and proposed an algorithm, based on learning control, to improve vehicle behavior and driver satisfaction. Our approach guarantees optimization of vehicle performance and adaptation to the driver´s style with road condition sensitivity. This improvement has been achieved by combining three knowledge acquisition sources: embedded dynamic models of powertrain, inductive inspection of driver actions and AT designer expertise; and by adding learning capabilities in order to significantly increase the system autonomy. Technically, GPS raises the following four problems which this paper addresses: (1) To achieve vehicle performance optimization of multiple antagonistic criteria, locally and globally over time, we considered a parametric disciminant function depending on an evaluation of the driver satisfaction and so called driver-style-state functions, as a reward for the system, and applied a reinforcement learning algorithm, derived from Q-learning method and combined with a mechanism to escape local optima. (2) Learning directly from the driver is performed when he selects AT ratio in manual mode. (3) Each driver´s personal style is represented by a Glass creation/selection mechanism. (4) GPS raises a few singularities which are addressed by a set of restriction rules derived from AT control expertise.