Compression ignition engine technologies have been advanced in the last decade to provide superior fuel economy and high performance. These technologies offer increased opportunities for optimizing engine calibration. Current engine calibration relies on deriving static tabular relationships between a set of steady-state operating points and the corresponding optimal values of the controllable variables. The values of these tabular relationships are interpolated to provide values of the controllable variables for each operating point while the engine is running. These values are controlled by the electronic control unit to achieve desirable engine behavior, for example in fuel economy, pollutant emissions, and engine acceleration performance. These strategies, however, are less effective during transient operation. Use of learning algorithms is an alternative approach that treats the engine as an "autonomous" system, namely, one capable of learning its optimal calibration for both steady-state and transient operating points in real time. In this approach, while the engine is running the vehicle, it progressively perceives the driver's driving style and eventually learns to operate in a manner that optimizes specified performance indices. Major challenges to this approach are problem dimensionality and learning time. This paper examines real-time, self-learning calibration of a diesel engine with respect to two controllable variables, i.e., injection timing and VGT vane position, to minimize fuel consumption.