A formal theory for the development of a generic model of an autonomous sensor is described. This model can be viewed as addressing a somewhat specific area of autonomous agent models. Perhaps the most intriguing aspect of this work is that it concentrates on the sensor itself whereas autonomous agents assume that the sensory input is complete and accurate to the degree that it needs to be. Building intelligent systems in engineering implies embedding the capabilities of the system operator in the model. These capabilities are not simply rules and facts, but also reasoning and decision making methodologies. So we present the notion of considering a sensor as an autonomous agent itself and call it highly autonomous, since no system can be completely autonomous. A highly autonomous sensor (HAS) not only interprets the acquired data in accordance with an embedded expert system knowledge base, but is also capable of learning and thereby improving its performance over time. This paper will concentrate on the learning aspects of the HAS model. The model is generic and can be used to instantiate any sensor as a HAS.