In the present article, an improved Learning Classifier Systems (LCS) is proposed to control the balance of a moving unmanned bicycle. Significant characteristics of learning classifier systems is that they can learn through a set of system actions in the real world (similar to intelligent creatures) while no dynamic model of the system is needed. Contrary to studies reported in the literature where action domain of the controller is discrete and accordingly such controller cannot be used in real world applications, in the present study efficacy of the classifier system is enhanced by definition of continuous domain for the outputs, and then is used to control the balance of unmanned bicycle. A scheme based upon fuzzy membership functions is proposed which makes it possible for the domain of actions to be continuous. The proposed LCS features a dynamic reward assignment mechanism which is invented to cope with the bicycle’s delayed response due to its mass inertias. This allows the rapid calculation of the reward and hence enables the controller to be used in such real time applications as the balance control of unmanned vehicles. A standard 2 degree of freedom (2-DOF) bicycle model is employed to demonstrate the efficiency of the enhanced LCS. Simulation results show that the proposed classifier system outperforms traditional classifier system as well as some of the more common balance-control strategies reported in the literature.