Sensorimotor Intelligence and Habits in Simulated Artificial Agents

Abstract

This thesis presents a set of simulated models to illustrate and interrogate notions of sensorimotor intelligence and habit from an enactive perspective. I draw upon methods from evolutionary robotics and developmental robotics to develop three new computational models concerning how certain sensorimotor structures can be viewed as self-organising and self-maintaining entities (i.e. habits). I use these models as robot controllers to explore how they produce concrete examples of adaptive behaviour in artificial agents, and I interpret these examples in terms of how they relate to our current understanding of natural cognition. Each of these models is intended as a contribution in its own right to future simulation efforts, but the primary focus in this thesis is their application to four specific investigations. These investigations have two primary subjects: Firstly, the role of sensorimotor contingencies (i.e. regularly recurring sensorimotor dynamics) in simple forms of cognition. In this, I find that simple sensorimotor contingencies are able to explain the capacity for certain cognitive capacities without the need to appeal to internal representations of the world, or even internal neural dynamics. Secondly, I ask how the habitual organisation of sensorimotor contingencies relates to adaptive and functional behaviour. Here I find that habit-based models can produce adaptive behaviour which is guided by the maintenance of internal sensorimotor structures, as opposed to specific functional outcomes of the behaviour. This brings us towards an understanding of the role of habit and sensorimotor intelligence in producing goal-directed behaviour, laying the groundwork for behaviour which is intrinsically meaningful for the agent that performs it.