Abstract: | Models of emotion are often suggested as a way of providing an evaluation of the current behaviour of an agent. In this work, we investigate whether emotions can actually provide suitable reinforcement signals for a Q-learning system to learn adaptive policies. For this purpose a recurrent network model of emotion consistent with the somatic marker hypothesis of Damasio was developed. Experimental work was done in a realistic mobile robot simulator in a simple foraging-like task. Experiments revealed that having emotions providing a context evaluation for direct use as a reinforcement signal does not work, but using them as modifiers for learning system parameters could be fruitful.
|