In order to discriminate among the different objects in its environment, an agent may develop a primitive notion of concepts based on the sensor data it receives. In this paper, this phenomenon is investigated by having software agents play discrimination games with the sensor data of autonomous robots. We have compared the Simple Prototype method and the Adaptive Subspace method. Both methods achieve high discrimination-success rates. The Adaptive Sub-space method accomplishes this with a converging and relatively low number of categories. The purpose of these discrimination games is to serve as a basis for lexicon formation experiments. From the experiments in this paper, we conclude that the Adaptive Subspace method is more attractive for discrimination games.