The relevance of decision making in autonomous systems is intrinsically related to the system capacity to discriminate its perception-action states. This is particularly challenging in unknown and changing complex environments, where providing a complete a priori representation to the system is not possible. To illustrate the problem, let us consider a decentralized control of road traffic, where a control device of the distributed infrastructure locally controls traffic, by learning to construct a precise representation (perception-action states) of the traffic state. In this context, it is challenging to define from prior knowledge a relevant representation of the traffic state that enables an efficient recommendation-based control. Without considering a prior domain-knowledge representation, we propose an approach able to combine a set of existing traditional unsupervised learning methods that collaborate as a population of agents in order to build an efficient representation. Our approach follows a constructivist learning perspective, where each agent produces a possible discretization of the raw sensed data. Thanks to a multi-agent reinforcement learning process, the population is able to collectively build a representation that combines the good capacities of the individual ones.