Special Issue on the Dynamical Systems Approach to Cognition

Perhaps the deepest philosophical insight of research in cognitive science and new A.I. in the last twenty years has been the realization that cognition is not “something done by the brain.” According to this view, cognitive activity must be understood as the result of interactions occurring within a set of coupled sub-systems belonging both to the agent and to the environment, inextricably intertwined, co-evolving and coadapting in real time. It is potentially misleading and often counter-productive to define features observable in the system’s behavior functionally as a property of the agent’s internal dynamics, its body morphology or the environment it is in; a canonic example is the coordination of the legs’ motion in insect-robots (Beer, 1995) whose architectures have no coordinating center and which critically depend on the local interaction between the leg morphodynamics and environment. At around the time this shift was taking place, research on neural networks was undergoing the boost that became known as “connectionism”. Although classic connectionism drew on the old computational metaphors about cognition, and technically aimed at a form of computation, a trend quickly developed to explore neural networks as dynamical systems, for example networks incorporating real time dynamics, as opposed to input-output timeless devices. The old strategy for constructing robots or understanding cognition presupposed a previous representation by an agent of the exterior world, its mapping and the construction of “internal” models of the world (for a critique of this view see Harvey, 1996 or Brooks, 1991). So, cognitive activity was seen as the control of physical action or physical properties by an informationtheoretic centered system and its specification was to be found at the algorithmic level. Only as far as physical properties could be constrained and dominated by the algorithmic level would one be able to deal with “interesting” cognitive behavior. The environment was treated as on independent provider of information input. For the new approach, cognitive activity arises from the interaction between the agent’s internal dynamics, its body morphodynamical properties and the history of interactions with its environment. From this point of view, roboticists and cognitive scientists reject, as Varela put it, that an agent’s cognitive activity can be described as aiming at a