Towards pro-active embodied agents: On the importance of neural mechanisms suitable to process time information

Abstract

In Embodied Cognitive Science, many studies have focused on reactive agents, i.e. agents that have no internal state and always respond in the same way to the same stimulus. However, this particular focus is not due to a rejection of the importance of internal states. Rather, it is due to the difficulty of developing pro-active embodied and situated agents, that is agents able to: (a) extract internal states by integrating sensorymotor information through time and, (b) later use these internal states to modulate their motor behaviour according to the current environmental circumstances. In this chapter we will focus on how pro-active agents can be developed and, more specifically, on which are the neural mechanisms that might favour the development of pro-active agents. By comparing the results of five sets of evolutionary experiments in which simulated robots are provided with different types of recurrent neural networks, we gain insight into the relation between the robots' capabilities and the characteristics of their neural controllers. We show how special mechanisms for processing information in time facilitate the exploitation of internal states. © 2006 Springer.