This paper addresses the use of emotions on autonomous agents for behaviour-selection learning,
focusing in the emotions fear, happiness and sadness. The control architecture is based in a motiva-
tional model, which performs homeostatic control of the internal state of the agent. The behaviour-
selection is learned by the agent using a Q-learning algorithm while there is no interaction with
other agents. In situations where interaction arises (e.g. interacting with other agents), agents rely on
stochastic games approaches as a learning strategy. The agent is intrinsically motivated and his final
goal is to maximize Happiness. The learning algorithms use happiness/sadness of the agent as posi-
tive/negative reinforcement signals. Fear is used to prevent the agent choosing dangerous actions or
being in dangerous states where non-controlled exogenous events, produced by external objects or
other agents, could danger him. Preliminary tests have been carried out in a virtual world, based in a
role-playing game.