Indre Pileckyte, Martin Takáč:  Computational model of intrinsic and extrinsic motivation for decision making and action-selection

Abstract:  This report describes a preliminary proposal for a computational model of decision control. Decision control is driven by needs of a simulated agent. The main novel aspect of the model is a combination of physiological/extrinsic needs (nutritional energy, physical energy, physical integrity) with psychological/intrinsic needs (curiosity, competence, playfulness). The current state of needs is represented as a point in a multidimensional space with needs as its dimensions – so called internal space. The agent performs actions, which have consequences in its environment (changes of external state) and also cause change of state in the internal space, however, the reward is determined purely by the effect on need satisfaction: The agent gets a positive reward if it gets closer to optimum in the internal state and a negative reward if it gets further from it. The action selection and decision control is driven by reinforcement learning. We combine model-free and model-based approaches. The report also contains simulation scenarios for testing the model.