This article focusses on the automated synthesis of agents in an uncertain environment, working in the setting of Reinforcement Learning, and more precisely of Partially Observable Markov Decision Processes. The agents (with no model of their environment and no short-term memory) are facing multiple motivations/goals simultaneously, a problem related to the field of Action Selection. We propose and evaluate various Action Selection architectures. They all combine already known basic behaviors in an adaptive manner, by learning the tuning of the combination, so as to maximize the agent's payoff. %This work opens the way to a second step in which the basic behaviors themselves will be selected and designed in an automated manner. The logical continuation of this work is to automate the selection and design of the basic behaviors themselves.