This is the abstract of a talk prepared for the International interdisciplinary seminar on new robotics, evolution and embodied cognition (IISREEC).12th to 15th November 2002, Lisbon, Portugal
Abstract: Does on-line intelligence (the production of fluid and adaptable responses to ongoing sensory input) constitute a hostile region of cognitive space where traditional explanatory concepts, such as representation and computation, find it hard to survive? If so, will dynamical systems theory inherit the cognitive earth? In the 1990s, as behaviour-based and evolutionary robotics gathered speed, it seemed to some of us that both of these questions should be answered in the affirmative. More recently, however, a certain rapprochement appears to have been on the cards. This is a good time, then, to take another look at the issues and arguments.
Following some conceptual clarifications, I focus on two threats to the idea that on- line intelligence must or should be explained by appeal to neurally located representations. These threats are, I explain, posed most sharply by evolutionary robotics. The first - the threat from non-trivial causal spread - occurs when extra- neural factors account for the kind of behavioural richness and flexibility normally associated with representation-based control. I show how this particular anti- representational challenge can, in principle, be met, if we apply the thought that, to be a representational system, an action-oriented neural system must not only be the source of at least some of the observed behavioural richness and flexibility, it must also feature two architectural traits, namely arbitrariness and homuncularity. This solution, however, opens the door to our second threat to representation - the threat from continuous reciprocal causation. In practice, the homuncularity condition will fail to be met by any system in which the causal contribution of each component is massively context-sensitive and variable over time. One of the lessons of evolutionary robotics is that the question of whether or not biological nervous systems exhibit dangerous levels of continuous reciprocal causation is nowhere near as clear cut as we might naively imagine.
Having found a (perhaps vulnerable) place for representations in the cognitive science of on-line intelligence, I argue that the same trick cannot be turned for a theoretically disciplined notion of computation, and that the dual phenomena of non- trivial causal spread and continuous reciprocal causation speak loudly in favour of a dynamical systems approach.